Radiosity and

Realistic Image
Synthesis

Michael F. Cohen
John R. Wallace

Academic Press Professional

Boston San Diago New York
London Syndey Tokyo Toronto
Copyright (c) 1995 by Academic Press, Inc.
GRAPHICS GEMS copyright (c) 1990 by Academic Press, Inc.
GRAPHICS GEMS II copyright (c) 1991 by Academic Press, Inc.
GRAPHICS GEMS III copyright (c) 1992 by Academic Press, Inc.
QUICK REFERENCE TO COMPUTER GRAPHICS TERMS
copyright (c) 1993 by Academic Press, Inc.
RADIOSITY AND REALISTIC IMAGE SYNTHESIS
copyright (c) 1993 by Academic Press Inc.
VIRTUAL REALITY APPLICATIONS AND EXPLORATIONS
copyright (c) 1993 by Academic Press Inc.

All rights reserved.

No part of this product may be reproduced or transmitted in any form or by any
means, electronic or mechanical, including input into or storage in any information
system, other than for uses specified in the License Agreement, without permission
in writing from the publisher.

Except where credited to another source, the C and C++ Code may be used freely to
modify or create programs that are for personal use or commercial distribution.

Produced in the United States of America

ISBN 0-12-059756-X
About the cover image:
The cover image shows the interior of Le Corbusiers Chapel at Ronchamp,
France. The illumination was computed using radiosity, with the sunbeams added
by stochastic ray tracing during rendering [109, 110]. The model was created by
Paul Boudreau, Keith Howie, and Eric Haines at 3D/EYE, Inc. with Hewlett-
Packards ARTCore Radiosity and Ray Tracing library.
The image is a frame from the animation The Key is Light presented at the
Siggraph 91 Electronic Theater. The video was produced by Hewlett-Packard
Company TV, with extensive help from Becky Naqvi, John Fujii, and Ron Firooz
at Hewlett-Packard Company.
The back cover image is a radiosity rendering from a scene of Luthers Tavern
in the Opera Tales of Hoffman. The opera lighting design software used for this
image is part of a PhD dissertation by Julie OBrien Dorsey at Cornell Universitys
Program of Computer Graphics [73].

Radiosity and Realistic Image Synthesis iii

Edited by Michael F. Cohen and John R. Wallace
 (a) (b)

(c) (d)

(e) (f)

Plate 1. Six Renderings of Red-Blue Box (see Chapter 1). (a) Local, (b) Ray
Trace, (c) Radiosity, (d) Radiosity + Glossy, (e) Radiosity + Fog, (f) Monte Carlo.
Courtesy of Michael Cohen, Holly Rushmeier, and Ben Trumbore, Program of
Computer Graphics, Cornell University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 2. A sculpture by
John Ferren entitled
Construction in Wood, A
Daylight Experiment.
Front faces of the panels
are white. The color is
caused by daylight
reflected from rear-facing
colored surfaces.
Courtesy of Cindy Goral,
Program of Computer
Graphics, Cornell
University.

Plate 3. A ray traced

image of the above
sculpture. All the panels
appear white since a
standard ray tracer cannot
simulate the
interreflection of light
between diffuse surfaces.
Courtesy of Cindy Goral,
Program of Computer
Graphics, Cornell
University.

Plate 4. A radiosity image

of the above sculpture.
Note the color bleeding
from the backs of the
boards to the fronts.
Courtesy of Cindy Goral,
Program of Computer
Graphics, Cornell
University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 6. Observer viewing images
projected onto frosted glass in
portrait cameras. Courtesy of Gary
Meyer, Program of Computer
Graphics, Cornell University.

Plate 5. Experimental setup to test Plate 7. Upside down views as seen

accuracy of radiosity method and by observer. Courtesy of Gary Meyer,
choice of color spaces. Courtesy of Program of Computer Graphics,
Gary Meyer, Program of Computer Cornell University.
Graphics, Cornell University.

Plate 8. Photograph of real scene Plate 9. Photograph of CRT screen

taken with portrait camera. (Color containing radiosity image. Courtesy of
adjusted for film and monitor Gary Meyer, Program of Computer
gamuts in Plates 8 and 9.) Courtesy Graphics, Cornell University.
of Gary Meyer, Program of Com-
puter Graphics, Cornell University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 10. Magritte
Studio. Radiosity with
texture mapping of both
reflecting surfaces and
light sources. Courtesy
of Michael Cohen,
Program of Computer
Graphics, Cornell
University.

Plate 11. Computer

Room. Shading using
direct illumination only.
Courtesy of Tamoyuki
Nishita, Fukuyama
University.

Plate 12. Auditorium.

An element mesh in
which T vertices have
been eliminated by
triangulation to create
conforming elements.
Courtesy of Daniel
Baum, Silicon Graphics
Corporation.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 13. Magritte
Studio, Lights Off.
Image created using the
same form factors as
plate 10. Turning off
light requires only
resolving the matrix
equation with new
emission values.
Courtesy of Michael
Cohen, Program of
Computer Graphics,
Cornell University.

Plate 14. Computer

Room. The same
environment as in Plate
11, with radiosity used
to compute both direct
and indirect illumina-
tion. Note the addi-
tional illumination on
the ceiling. Courtesy of
Tamoyuki Nishita,

Plate 15. The same

image as in Plate 12
with out displaying the
mesh. Courtesy of
Daniel Baum, Silicon
Graphics Corporation.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
Plate 16. Steel Mill. A complex environment shaded using progressive refine-
ment radiosity. Courtesy of John Wallace and Stuart Feldman, Program of
Computer Graphics, Cornell University.

Plate 17. Constuctivist Museum. The complex interreflection from the ceiling
baffles was simulated with the progressive refinement approach. Courtesy of
Shenchang Chen, Stuart Feldman, and Julie OBrien Dorsey, Program of Com-
puter Graphics, Cornell University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 18. Plate 19.

Plate 20. Plate 21.

A Sequence showing the links formed at each level of a hierarchy generated by

Hanrahan, Salzman, and Aupperles algorithm. Courtesy of Pat Hanrahan,
Princeton University.

Plate 22. Final image with

texture mapping. Courtesy of
Pat Hanrahan, Princeton
University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 23. Radiosity
solution. Courtesy of
Brian Smits, James Arvo,
and David Salesin,
Program of Computer
Graphics, Cornell
University.

Plate 24. Importance

solution. Courtesy of
Brian Smits, James Arvo,
and David Salesin,
Program of Computer
Graphics, Cornell
University.

Plate 25. Combined

radiosity and importance
solutions. Courtesy of
Brian Smits, James Arvo,
and David Salesin,
Program of Computer
Graphics, Cornell
University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
Plate 26. Radiosity/Importance Plate 27. Radiosity/Importance
solution with mesh. Courtesy of solution after reconstruction.
Brian Smits, James Arvo, and David Courtesy of Brian Smits, James
Salesin, Program of Computer Arvo, and David Salesin, Program
Graphics, Cornell University. of Computer Graphics, Cornell
University.

Plate 28. Radiosity solution from Plate 29. Importance solution.

further back. Courtesy of Brian Courtesy of Brian Smits, James
Smits, James Arvo, and David Arvo, and David Salesin, Program
Salesin, Program of Computer of Computer Graphics, Cornell
Graphics, Cornell University. University.

Plate 30. Radiosity from even Plate 31. Importance from even
further back. Courtesy of Brian further back. Courtesy of Brian
Smits, James Arvo, and David Smits, James Arvo, and David
Salesin, Program of Computer Salesin, Program of Computer
Graphics, Cornell University. Graphics, Cornell University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 32. Radiosity
solution using quadtree
based adaptive subdivi-
sion. Failure to resolve
discontinuities results in
the inaccurate representa-
tion of shadow bound-
aries. Courtesy of Filippo
Tampieri and Dani
Lischinski, Program of
Computer Graphics,
Cornell University.

Plate 33. Radiosity

solution of same environ-
ment as above, but with
the use of discontinuity
meshing. Courtesy of
Filippo Tamieri and Dani
Lischinski, Program of
Computer Graphics,
Cornell University.

Plate 34. Use of disconti-

nuity meshing to create
accurate shadow bound-
aries. Courtesy of Filippo
Tamieri and Dani
Lischinski, Program of
Computer Graphics,
Cornell University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 35. Multipass solution
after the initial progressive
radiosity solution. Total time:
approx. 12 minutes. Courtesy of
Shenchuang Chen, Apple
Computer Corporation.

Plate 36. Multipass solution:

Direct illumination computed
with Monte Carlo ray tracing,
caustics computed with light ray
tracing, combined with indirect
component of initial progressive
radiosity solution. Total time:
approx. 4.5 hours Courtesy of
Shenchuang Chen, Apple
Computer Corporation.

Plate 37. Components of Plate 36. Direct Monte Carlo + Indirect Progressive
Refinement Radiosity + Light Ray Tracing. Courtesy of Shenchuang Chen,
Apple Computer Corporation.

Plate 38. Multipass solution

after full Monte Carlo solution
for both direct and indirect
illumination. Total time: approx
21 hours. Courtesy of
Shenchuang Chen, Apple
Computer Corporation.

Plate 39. Components of Plate 38. Direct + Indirect Monte Carlo + Light Ray
Tracing. Courtesy of Shenchuang Chen, Apple Computer Corporation.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 40. A ships boiler
room, with Phong
highlights added to a
progressive radiosity
solution during rendering.
Courtesy of John Wallace,
John Lin, and Eric
Haines, Hewlett-Packard
Corporation.

Plate 41. Radiosity

solution for indirect
illumination, with the
direct illumination
computed at each pixel
during rendering. Bump
mapping is performed
during the per-pixel
illumination computation.
Courtesy of Peter Shirley.

Plate 42. Bidirectional

ray tracing. The caustic
on the table is caused by
light focused through the
glass and was computed
using light ray tracing.
Courtesy of Peter Shirley.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
Plate 43. Radiosity solution without Plate 44. Radiosity solution with
inclusion of specular to diffuse extended form factors to capture
reflection of light off mirror. light reflected from mirror. Courtesy
Courtesy of Franois Sillion, Ecle of Franois Sillion, Ecle Normale
Normale Supriuere. Supriuere.

Plate 45. Dutch

Interior, after
Vermeer. A two-
pass solution:
radiosity plus the
reflection frustum
algorithm during
rendering to com-
pute glossy reflec-
tion from floor to
eye. Courtesy of
John Wallace,
Program of Com-
puter Graphic,
Cornell University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 46. Computation of glossy
and mirror specular reflection
using spherical harmonics to
approximate directional radiance
distribution. Courtesy of Franois
Sillion, Program of Computer
Graphics, Cornell University.

Plate 47. Main Council

chamber in the new
Jerusalem City Hall.
Designed by A. J.
Diamond, Donald Schmitt
and Co. Rendered using
radiosity software being
developed at Lightscape
Graphics. Courtesy of
Stuart Feldman,
Lightscape Graphics
Software.

Plate 48. Use of

zonal method to
include a participating
medium (smoke)
within a radiosity
solution. Courtesy of
Holly Rushmeier,
Program of Computer
Graphics, Cornell
University.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 49. A unified solution
for Lambertian diffuse,
glossy, and mirror specular
reflection using spherical
harmonics to approximate
radiance distribution.
Courtesy of Franois
Sillion, Program of Com-
puter Graphics, Cornell
University.

Plate 50. The main

council chamber in
Plate 47. Courtesy of
Stuart Feldman,
Lightscape Graphics
Software.

Plate 51.
Gemldegalerie
BERLIN. Image
produced using the
COPHOS lighting design
software under develop-
ment at Zumtobel Licht
GmbH. Courtesy of
Zumtobel GmbH, Austria.

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
 Plate 52. Home of the
Brain, from a project on
Virtual Reality and
Telecommunications.
Courtesy of Monika
Fleischmann and
Wolfgang Strauss,
ART+COM, Berlin

Plate 53. Scene from the

opera Turandot, rendered
with software for stage
lighting design. Courtesy of
Julie OBrien Dorsey,
Program of Computer
Graphics, Cornell Univer-
sity.

Plate 54. Scene of

Venice from Tales of
Hoffman. Courtesy of
Julie OBrien Dorsey,
Program of Computer
Graphics, Cornell

Radiosity and Realistic Image Synthesis

Edited by Michael F. Cohen and John R. Wallace
Contents

Foreword by Donald Greenberg xi

Preface xiii

1 Introduction 1
1.1 Realistic Image Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 A Short Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Raster Graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2 Global Illumination Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Early Radiosity Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.4 The Rendering Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Radiosity and Finite Element Methods . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 The Radiosity Method and This Book . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Rendering Concepts by Pat Hanrahan 13

2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Basic Optics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Radiometry and Photometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 The Light Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.1 Transport Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.2 Radiance and Luminance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4.3 Irradiance and Illuminance . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.4 Radiosity and Luminosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.5 Radiant and Luminous Intensity . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.6 Summary of Radiometric and Photometric Quantities . . . . . . 27
2.5 Reflection Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5.1 The Bidirectional Reflection distribution Function . . . . . . . . 28
2.5.2 Mirror Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5.3 The Reflectance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.4 Lambertian Diffuse Reflection . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5.5 Glossy Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6 The Rendering Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.6.1 Local or Direct Illumination . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Radiosity and Realistic Image Synthesis iv

Edited by Michael F. Cohen and John R. Wallace
 2.6.2 Global or Indirect Illumination . . . . . . . . . . . . . . . . . . . . . . . . 38
2.6.3 The Radiosity Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Discretizing the Radiosity Equation 41

3.1 The Radiosity Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Making Image Synthesis Tractable . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 The Radiosity Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Approximating Radiosity across a Surface . . . . . . . . . . . . . . . . . . . . 48
3.5 Error Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5.1 Point Collocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5.2 Galerkin Form of Weighted Residuals . . . . . . . . . . . . . . . . . . 56
3.6 Constant Element Radiosities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Higher-order Basis Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.8 Parametric Mapping to a Master Element . . . . . . . . . . . . . . . . . . . . . 61
3.8.1 Master Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.8.2 Isoparametric Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 The Form Factor 65

I. The Form Factor Integral 65
4.1 The Coefficients of K . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2 The Differential Form Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Three Formulations of the Form Factor . . . . . . . . . . . . . . . . . . . . . . . 69
4.4 Computing the Form Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
II. Closed Form Solutions for the Form Factor 72
4.5 Formulae for Simple Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.6 Differential Area to Convex Polygon . . . . . . . . . . . . . . . . . . . . . . . . 72
4.7 General Polygon to Polygon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
III. Numerical Solutions for the Form Factor 75
4.8 Numerical Integration in General . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.8.1 Gaussian Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.8.2 Quadrature Points and the Form Factor Integral . . . . . . . . . . 77
4.8.3 Monte Carlo Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.9 Evaluating the Inner Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.9.1 Hemisphere Sampling Algorithms . . . . . . . . . . . . . . . . . . . . . 79
4.9.2 Nusselt Analog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.9.3 The Hemicube . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.9.4 Single-Plane Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.9.5 Monte Carlo Ray Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.9.6 Area Sampling Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.10 Full Area-to-Area Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

Radiosity and Realistic Image Synthesis v

Edited by Michael F. Cohen and John R. Wallace
 4.10.1 Monte Carlo Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.11 Contour Integral Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.12 A Simple Test Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.13 Nonconstant Basis Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.13.1 The Hemicube for General Form Factors . . . . . . . . . . . . . . . . 99
4.13.2 Monte Carlo for General Form Factors . . . . . . . . . . . . . . . . . 99
4.13.3 Singularities in the Integrand . . . . . . . . . . . . . . . . . . . . . . . . 100
4.14 Acceleration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.14.1 Hemicube Acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.14.2 Ray Tracing Acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

5 Radiosity Matrix Solutions 109

5.1 Qualities of the Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.2 Linear System Solution Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.2.1 Direct Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.2.2 Iterative Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3 Relaxation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.3.1 Jacobi iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.3.2 Gauss-Seidel Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.3.3 Southwell Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.3.4 Ambient Energy and Overelaxation . . . . . . . . . . . . . . . . . . . 122
5.4 Dynamic Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.4.1 Lighting Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.4.2 Reflectivity Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.4.3 Changes in Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.5 Parallel Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

6 Domain Subdivision 131

6.1 Error Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.1.1 True Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.1.2 Local Estimate of Approximation Error . . . . . . . . . . . . . . . . 132
6.1.3 Residual of the Approximate Solution . . . . . . . . . . . . . . . . . 134
6.1.4 Error Based on the Behavior of the Kernel . . . . . . . . . . . . . 135
6.1.5 Image Based Error Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.1.6 Perceptually Based Error Metrics . . . . . . . . . . . . . . . . . . . . . 136
6.2 Mesh Characteristics and Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.2.1 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.2.2 Mesh Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.2.3 Element Order and Continuity . . . . . . . . . . . . . . . . . . . . . . . 142
6.2.4 Element Shape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.2.5 Discontinuities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.3 Automatic Meshing Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

Radiosity and Realistic Image Synthesis vi

Edited by Michael F. Cohen and John R. Wallace
 6.3.1 A Posteriori Meshing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.3.2 Adaptive Subdivision: H-refinement for Radiosity . . . . . . . 157
6.3.3 Error Estimation for Adaptive Subdivision . . . . . . . . . . . . . 159
6.3.4 Deciding How to Subdivide . . . . . . . . . . . . . . . . . . . . . . . . . 165

7 Hierarchical Methods 167

I. Hierarchical Subdivision 168
7.1 A Physical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.2 Two-Level Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.3 The K Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
7.4 Multilevel hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7.4.1 N-Body Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.4.2 Radiosity and the N-Body Problem . . . . . . . . . . . . . . . . . . . 177
7.4.3 Hierarchical Refinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.4.4 Solution of the Hierarchical System . . . . . . . . . . . . . . . . . . . 181
7.4.5 The Oracle Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.4.6 Progressive Refinement of the Hierarchy . . . . . . . . . . . . . . . 184
7.4.7 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
II. Hierarchical Basis Functions and Wavelets 187
7.5 Hierarchical Basis Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.6 Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.6.1 Haar Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.6.2 Vanishing Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.6.3 Vanishing Moments and Sparse Representations . . . . . . . . . 194
7.6.4 A Wavelet Radiosity Algorithm . . . . . . . . . . . . . . . . . . . . . . 198
III. Importance-Based Radiosity 201
7.7 Importance Meshing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
7.7.1 The Importance Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
7.7.2 Importance-Based Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.8 Hierarchical Radiosity and Importance . . . . . . . . . . . . . . . . . . . . . . 205
7.8.1 Pseudocode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
7.8.2 Example Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

8 Meshing 209
8.1 Basic Subdivision Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
8.2 Mesh Template Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
8.2.1 Grid Superposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
8.2.2 Template Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
8.2.3 Multiblocking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
8.2.4 Adaptive Subdivision with Templates . . . . . . . . . . . . . . . . . 214
8.3 Decomposition Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
8.3.1 Nodes-Elements-Together Decomposition . . . . . . . . . . . . . . 217

Radiosity and Realistic Image Synthesis vii

Edited by Michael F. Cohen and John R. Wallace
 8.3.2 Decomposition by Recursive Splitting . . . . . . . . . . . . . . . . . 217
8.3.3 Decomposition by Advancing Front . . . . . . . . . . . . . . . . . . . 218
8.3.4 Nodes-First Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 219
8.4 Mesh Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
8.5 Discontinuity Meshing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
8.5.1 Discontinuities in Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
8.5.2 First and Second Derivative Discontinuities . . . . . . . . . . . . 224
8.5.3 Shadow Volume Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 229
8.5.4 Critical Surface Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 231
8.6 Topological Data Structures and Operators . . . . . . . . . . . . . . . . . . . 234
8.6.1 Data Structure Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
8.6.2 The Winged-Edge Data Structure . . . . . . . . . . . . . . . . . . . . . 235
8.7 Alternatives to Meshing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

9 Rendering 243
9.1 Reconstructing the Radiosity Function . . . . . . . . . . . . . . . . . . . . . . 244
9.2 Interpolation Methods for Rendering . . . . . . . . . . . . . . . . . . . . . . . . 245
9.2.1 C0 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
9.2.2 C1 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
9.3 Two-Pass Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
9.3.1 Evaluating the Radiosity Equation per Pixel . . . . . . . . . . . . 259
9.3.2 Multi-Pass Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
9.4 Incorporating Surface Detail . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
9.4.1 Texture Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
9.4.2 Bump Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
9.5 Mapping Radiosities to Pixel Colors . . . . . . . . . . . . . . . . . . . . . . . . 267
9.5.1 Gamma Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
9.5.2 Real-World Luminance to Pixel Luminance . . . . . . . . . . . . 268
9.6 Color . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
9.6.1 Human Vision and Color . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
9.6.2 Color Matching Functions and the CIE Chromaticity Di-
agram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
9.6.3 Color Spaces and Image Synthesis . . . . . . . . . . . . . . . . . . . . 280
9.6.4 Direct Use of Spectral Data . . . . . . . . . . . . . . . . . . . . . . . . . 283
9.7 Hardware Accelerated Rendering . . . . . . . . . . . . . . . . . . . . . . . . . . 284
9.7.1 Walkthroughs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
9.7.2 Hardware-Supported Texture Mapping . . . . . . . . . . . . . . . . 285
9.7.3 Visibility Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286

10 Extensions 289
10.1 Nondiffuse Light Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
10.1.1 Form Factors to and from Light Sources . . . . . . . . . . . . . . . 290

Radiosity and Realistic Image Synthesis viii

Edited by Michael F. Cohen and John R. Wallace
 10.1.2 Point Lights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
10.1.3 Parallel Lights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
10.1.4 General Luminaires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
10.1.5 Spot Lights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
10.1.6 Sky Light . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
10.1.7 Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
10.1.8 Light Source Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
10.2 Directional Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
10.2.1 Classifying Transport Paths . . . . . . . . . . . . . . . . . . . . . . . . . 299
10.2.2 Tracing the Transport Paths . . . . . . . . . . . . . . . . . . . . . . . . . 302
10.2.3 Implicit Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
10.2.4 Explicit Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
10.2.5 Non-Lambertian Reflection and Hierarchical Methods . . . . 316
10.2.6 Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
10.2.7 Two-Pass Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
10.2.8 Surface Reflectance/Transmittance Data . . . . . . . . . . . . . . . 324
10.3 Participating Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
10.3.1 Path Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
10.3.2 The Zonal Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

11 Applications and Research 331

11.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
11.1.1 Architectural Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
11.1.2 Lighting Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
11.1.3 Remote Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
11.1.4 Visual Shape Understanding . . . . . . . . . . . . . . . . . . . . . . . . 338
11.1.5 Infrared Signature Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 339
11.1.6 Fine Arts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
11.2 Experimental Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
11.3 Future Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
11.3.1 Error Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
11.3.2 Perceptually Based Error Metrics . . . . . . . . . . . . . . . . . . . . . 343
11.3.3 Physically Based Emission and BRDF Data . . . . . . . . . . . . 344
11.3.4 Meshing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
11.3.5 Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
11.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347

Bibliography 349

Index 373

Radiosity and Realistic Image Synthesis ix

Edited by Michael F. Cohen and John R. Wallace
Foreword
For the past 25 years, researchers in the field of computer graphics have
continuously striven for the production of realistic images of nonexistent envi-
ronments. To attain this goal and its ultimate potential for design and aesthetic
evaluations, it is necessary to accurately represent the appearance of objects and
scenes as they look to us. This requires the knowledge of how to simulate both
the physical behavior of light and the perceptual behavior of the human visual
system.
The accurate simulation of physical processes is crucial for realistic image
synthesis. Ad hoc procedures, despite the fact that they can produce pretty
pictures, will not suffice. The radiosity method, originally based on principles
of thermodynamics, provides this physical basis and establishes the foundations
for future rendering and display systems.
More explicitly, the creation of photorealistic images requires four basic
components, a local model of light reflection, a means for simulating the propa-
gation of energy throughout an environment, the appropriate strategies for sam-
pling the scene, and procedurally accurate methods for displaying the results.
The radiosity method discussed in this book describes each of these steps in
great detail.
Historically, a major argument against the use of radiosity procedures has
been the excessive computing demands. Today these constraints are rapidly
being eliminated. During the last decade alone, processing power of workstations
and personal computers has increased by three orders of magnitude. However
skeptical one might be, all indications are that the trend of almost doubling
computer power each year will continue until at least the end of this decade.
Memory and storage costs have also dropped, by approximately four orders
of magnitude since the early 1970s. Most recently, new advances in network
technology have improved the possibility for image transmission rates by six
orders of magnitude from what was available two decades ago. Further advances
in the technology will occur due to parallelism and compression schemes.
Display technology is also accelerating at a remarkable pace. The dot spac-
ing in printing technologies has been vastly reduced. High-resolution display
monitors are now commonplace. The advent of high-definition television will
push video technology further, both in terms of refresh rates and display res-
olution, and ultimately in cost due to the economics of mass production. For
normal viewing conditions, resolutions will have surpassed the visual acuity of
the human eye. Intensity ranges will be increased, and the speed of displays is
already sufficiently fast to imply continuous motion.
With these dramatic advances in computing and display technologies, the

Radiosity and Realistic Image Synthesis x

Edited by Michael F. Cohen and John R. Wallace
FORWARD

arguments against the computational complexity of image synthesis techniques

fall hollow. Processing and storage will essentially be free, and transmission
will be sufficiently fast to deliver high quality picture information and allow the
use of remote computing nodes. The computing obstacles of the past will have
been overcome.
What is now needed is the ability to mimic the complex physical behavior
of light distribution, from microscopic to macroscopic ranges. The radiosity
method for image synthesis provides the theoretical underpinnings and algorith-
mic techniques toward these ends. With future experimental measurements and
comparisons, these methods can be continually refined to improve their accuracy.
This book is the most thorough treatise on the radiosity method yet to be
published in the field of computer graphics. The text includes detailed descrip-
tions of all of the major components required to create a system for displaying
modeled environments. From the explanations of the fundamental scientific
bases to the state-of-the-art algorithms for implementation, the topics are cov-
ered in a clear and comprehensive way. The authors are to be congratulated
for their in-depth treatment of the subject and for the presentation of a text
that can significantly influence rendering systems of the future. The quest for
photorealism will continue!

Donald P. Greenberg
Professor and Director
Program of Computer Graphics
Cornell University

Radiosity and Realistic Image Synthesis xi

Edited by Michael F. Cohen and John R. Wallace
Preface
Over the past decade, computer graphics has exploded out of university re-
search laboratories onto television and cinema screens, and into medical imag-
ing, scientific visualization and computer-aided design systems. A persistent
goal through much of the research that has contributed to these developments
has been to recreate, with the computer, strikingly realistic images of environ-
ments that do not (and often could not) exist. This field of endeavor has come
to be known as realistic image synthesis. Radiosity provides one important ap-
proach to evaluating a physically-based illumination model, which is a key part
of image synthesis.
The number of papers published on radiosity and related techniques increases
yearly. Although the field is by no means mature, it is at a transition point, with
early intuitive methods being replaced by approaches based on more rigorous
attention to underlying physical processes and numerical methods. Thus, this is
a natural time to summarize the research to date and to present it in a uniform
format.
Our goal in writing this book is to survey the state-of-the-art in radiosity
and related image synthesis research, to explain the underlying theory, and to
provide a framework that organizes the broad and growing literature surround-
ing this field. The book is intended for those interested in pursuing research in
global illumination and image synthesis. It should also provide a useful theoret-
ical background and insight into many practical issues, for those implementing
radiosity or other global illumination systems.
After a short introductory chapter, the book continues with a chapter by Pat
Hanrahan that carefully defines the terminology and concepts of radiometry and
photometry, the fields concerned with the measurement of light. This discussion
ends with the derivation of the rendering equation and its specialization in the
form of the radiosity integral equation. The following three chapters discuss
the use of finite element methods to solve this equation, by first formulating an
approximately equivalent set of linear equations, then evaluating the coefficients
of the linear system (the form factors), and finally solving the resulting matrix
equation.
This is followed by three chapters in which the topic of domain subdivision
(or meshing) is discussed. The discussion begins with an overview of mesh-
ing issues, then takes an aside to discuss new hierarchical formulations of the
radiosity problem including applications of wavelet methods, and closes with a
chapter on the practical issues in generating a good mesh.
Chapter 9 explores the final step in the image synthesis process, that is,
mapping the results of the numerical simulation to a display device. In this

Radiosity and Realistic Image Synthesis xii

Edited by Michael F. Cohen and John R. Wallace
PREFACE

context, the peculiarities of the human visual system are discussed, ranging
from the nonlinear response of the eye to luminance, to the tristimulus theory of
color perception. Chapter io then expands the scope of the radiosity methods by
lifting many of the restrictions assumed in the earlier discussion, such as diffuse
surfaces and non-participating media. Finally, the book concludes with a chapter
that explores a number of developing applications of the radiosity method, and
takes a moment to look towards the future.
The presentation in this book assumes a familiarity with the basic concepts
of computer graphics. There are a number of excellent computer graphics texts
that more fully explore some of the techniques that are called on in the algo-
rithms described here [84, 97, 173, 195, 258]. The discussion also assumes
an understanding of undergraduate calculus and linear algebra. Where more
advanced mathematical concepts are required, an effort is made to provide the
reader with enough background information to understand and appreciate the
material.

Acknowledgments
We thank the many colleagues who have directly and indirectly contributed
to the making of this book.
Without the dedication and persistent efforts of Prof. Donald P. Greenberg
of Cornell University, neither author would be in a position today to write this
text. His contributions to the development of the field of image synthesis are
well known. We thank him personally for inviting us into Cornells Program of
Computer Graphics where both authors were introduced to radiosity and image
synthesis, and for contributing the Foreword to this book.
Pat Hanrahan, beyond contributing a chapter to the book, is also largely
responsible for providing the first author with the stimulating environment at
Princeton University in which to work.
We would like to especially acknowledge the great efforts that went into
reviewing chapters of this book by Ken Chiu, Robert Cross, Brian Curless,
Stuart Feldman, Alain Fournier, John Fujii, Steven Gortler, Paul Lalonde, Marc
Levoy, Robert Lewis, Dani Lischinski, Earlin Lutz, Holly Rushmeier, David
Salesin, Peter Shirley, and Filippo Tampieri.
We thank Jutta Joesch for many hours of editing this text and for her enor-
mous help in gaining a better understanding of how to explain many of the more
difficult concepts presented. We would also like to thank Steven Gortler and
Peter Schrder for many discussions leading to much of the material on wavelets
in Chapter 7; Holly Rushmeier for numerous discussions that contributed ma-
terially to the content of this book; John Abel, Maged Tawfik, Paul Heckbert,
Mark Reichert, Seth Teller, David Munson, and Stuart Feldman for valuable

Radiosity and Realistic Image Synthesis xiii

Edited by Michael F. Cohen and John R. Wallace
PREFACE

discussions; John Fujii for first pointing out the topological shadow test dis-
cussed in Chapter 8, and for many hours of enjoyable discussions of aesthetic
and philosophical questions; Tamar Cohen for creating models used in some of
the images; Emil Ghinger for the black and white photography; Kevin Stokker
for software used to compute the error images in Chapter 6; Kim Wagner for
help in obtaining the cover image; Eric Haines for providing the initial version
of the Bibliography; Brian Rosen for help in compiling the Bibliography.
The authors would like to acknowledge some of the many additional collabo-
rators through the past decade who have contributed to this work. These include
Daniel Baum, Philip Brock, Rikk Carey, Shenchang Chen, Lisa Desjarlais, Stu-
art Feldman, Cindy Goral, Kevin Koestner, David Immel, Peter Kochevar, Alan
Polinsky, David Salmon, Kenneth Torrance, Ben Trumbore, and many others
at Cornell University; Franois Sillion and Claude Puech at the Ecle Normale
Suprieure, James Painter, John Kawai, and Gershon Elber at the University of
Utah, Philipp Slusallek at Universitt Erlangen, and many current colleagues at
Princeton University.
We would like to thank Eric Haines and Kells Elmquist at 3D/EYE, Inc. for
many years of collaboration in the pursuit of realistic image synthesis, Samir
Hanna for providing the second author time to write this all down, and the many
other people at 3D/EYE, Inc. and Hewlett-Packard who have jointly participated
in the development of radiosity and rendering software.
Images were contributed by Daniel Baum, A. T. Campbell III, Julie 0Brien
Dorsey, Shenchang Chen, Stuart Feldman, Monika Fleischmann, Cindy Goral,
Eric Haines, Pat Hanrahan, Paul Heckbert, Keith Johnson, Dani Lischinski, Gary
Meyer, David Munson, Mark Reichert, Holly Rushmeier, Brian Smits, David
Salesin, Peter Shirley, Franois Sillion, Filippo Tampieri, Hewlett Packard, and
Zumtobel Licht GmbH.
To Jenifer Niles, our editor at Academic Press, thank you for guiding us
successfully through the process of creating an actual book.
Finally, the contribution of our wives, Jutta M. Joesch and Diane L. Wallace
cannot be understated. Without their patience and support we could not have
finished this.

Michael F. Cohen John R. Wallace

Department of Computer Science 3D/EYE, Inc.
Princeton University Ithaca, NY

Radiosity and Realistic Image Synthesis xiv

Edited by Michael F. Cohen and John R. Wallace
 But something in the air sets me to thinking, there might be things
not too far off, that I might tell a few stories about, someday myself.
Though exactly how Ill do its beyond me. It wouldnt be any too
simple, just trying to describe this scene right here, how pretty a
figure that bird cuts, sailing across the red horizon. And l took
these sharp eyes to be a blessing. When they might, just as easily,
turn out to be a curse.

Oh well, enough of these idle musings. They aint gonna feed me.
Id better get down to business.

Alan Cohen
from The Saga of Harry the Snake

Radiosity and Realistic Image Synthesis xv

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION

Chapter 1

Introduction

In the pursuit of lifelike images, artists have long attempted to understand the
behavior of light and the characteristics of perception. Techniques that may
appear obvious, like perspective, were developed through painstaking study and
experimentation. The paintings of Vermeer and Rembrandt represent an under-
standing of illumination, color, and perception that evolved through centuries
of such experience. More recently, the Impressionists made a particular study
of the subtleties of light and shading; Renoir, for example, pointed out that
Shadows are not black; no shadow is black. It always has color.1
The connection between light and visual representation received its most
concrete realization with the invention of photography in the nineteenth century.
Because a photograph is the direct consequence of the physical propagation of
light, the camera is an invaluable recorder of things that exist. The creation of
realistic images of things that do not exist, or that are not normally perceivable
as images, such as scientific data, has remained until recently the domain of the
artist and illustrator.

1.1 Realistic Image Synthesis

Over the last few centuries physicists have developed mathematical models of
the processes by which light interacts with surfaces and propagates through an
environment. With the advent of the computer it has become practical to evaluate
such models on a large enough scale to simulate complex phenomena. Using
a computer, a model of light reflection and propagation can be evaluated for a
scene whose geometry and material properties have been specified numerically.
In effect, a photograph can be taken of a scene that does not exist in reality.
The ability to create images of nonexistent environments is important to ap-
plications ranging from industrial or architectural design to advertising and enter-
tainment. Phenomena not accessible to normal visual experience can also be vi-
1
The immediate source of this quotation, which comes close to reducing radiosity to a
sentence, is Parker et al. [179], who in turn quote from [193].

Radiosity and Realistic Image Synthesis 1

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.1 REALISTIC IMAGE SYNTHESIS

sualized by applying the illumination model to other forms of three-dimensional

data. For example, data from magnetic resonance imaging can be rendered to
provide three-dimensional images of the inside of the body.
The creation of images by evaluating a model of light propagation is called
image synthesis and has been studied extensively in the field of computer graph-
ics since the 1970s. The goal of image synthesis is often stated as photorealism.
However, although photography produces realistic images, it is a physical pro-
cess subject to the constraints of camera optics and the chemical nature of film.
Should image synthesis really attempt to simulate photography, or should it aim
higher?

1.1.1 Goals
A clear understanding of the goal of image synthesis becomes increasingly im-
portant as algorithms and computational methods grow more sophisticated. In
addition to the evaluation of competing approaches, more intelligent algorithms
need a basis for deciding how to allocate computational effort and when to end
the computation, which requires knowing when the goal has been achieved.
Perhaps the most far reaching goal for image synthesis is the creation a
visual experience identical to that which would be experienced in viewing the
real environment. The diagram in Figure 1.1 shows a simple model of the
image synthesis process that provides a basis for discussing the issues involved
in reaching this goal.
In the real world, as shown in the top half of the diagram, light propagates
through the scene and eventually enters the eye with a particular directional
and wavelength distribution. The eye and the brain process this information at
increasingly higher levels of abstraction, leading ultimately to what is called the
visual experience.
The bottom half of the diagram shows the modifications to the process
required for image synthesis. Instead of the physical propagation of light, a
mathematical model is evaluated to produce the required distribution of light
energy. These results are then passed to a display device that physically realizes
the computed light distribution and sends it to the eye. Image synthesis thus
appears to require simply the exact reproduction of the distribution of light
energy entering the eye. Given this, the process of experiencing the image will
take care of itself.

1.1.2 Limitations
There are two problems with this apparently simple approach. First, the com-
putation in step one is arbitrarily expensive. For all practical purposes, there is
no end to the detail or accuracy with which reality might be simulated. How

Radiosity and Realistic Image Synthesis 2

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.1 REALISTIC IMAGE SYNTHESIS

Figure 1.1: The process of visual experience. The top half of the figure dia-
grams real-world experience; the bottom half displays visual experience based
on computer simulation.

should limited computational resources be distributed? When is the simulation

done?
The second problem is with the display device. Even assuming that the
first step is performed perfectly, there is no existing device that can correctly
perform the second step! We can only imagine what such a device might be
likeperhaps a descendant of current virtual-reality goggles, with extremely
high spatial and color resolution, a field of view encompassing the entire range
of our peripheral vision, and the ability to reproduce luminances ranging from
starlight to the glare of snow on a sunny day.
In todays reality, the device will likely consist of a cathode ray tube (CRT),
which generates a two-dimensional map of discrete picture elements with a spa-
tial resolution of 1280 by 1024 pixels (often much less) and a color resolution
of 256 values for each of three color channels. The range, or gamut, of repro-
ducible colors will depend on the particular phosphors used in the CRT. Viewing
conditions, such as the ambient light level in the room containing the CRT, will
partially determine the eyes response to the light leaving the CRT. In most cases
a single image will be presented to both eyes.

Radiosity and Realistic Image Synthesis 3

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.2 A SHORT HISTORICAL PERSPECTIVE

In part because of the limitations of available devices, the goal of image

synthesis is, in practice, the reproduction of an image rather than of a direct
visual experience. This goal maps more directly to the currently available 2D
device (the CRT). The goal is similar but not identical to photorealism in that it
does not necessarily include reproducing all the characteristics of photography.
The limitations of the display device provide one set of guidelines for the
computation. For example, there is no point in computing a simulation with
a spatial or color resolution greater than that reproducible by the device. An
understanding of the final perceptual steps of the process is also important to
guiding the development of image synthesis algorithms. Based on an under-
standing of perception one can focus computational resources on aspects of the
simulation that contribute most to the final visual experience. For example,
the eye is particularly sensitive to contrast in luminance while being relatively
insensitive to absolute luminance levels.
The subject of this book is primarily the first part of the image synthesis
process, the computation of the light distribution at an image plane. This requires
developing a mathematical model of light propagation. The model may contain
certain simplifying assumptions; the radiosity method, for example, is initially
based on the assumption that all surfaces reflect light diffusely. Analytical or
numerical methods can then be developed to evaluate the mathematical model.
Algorithms that implement these solution methods must be written and, finally,
the results must be displayed as an image. These steps will form the basic
content of this book.
The evaluation of an illumination model cannot proceed until one has a
mathematical description of the environment to be rendered. The specification
of the scene geometry and material properties is itself a topic of active research
and presents many difficulties. This problem will not be addressed in this book.

1.2 A Short Historical Perspective

The radiosity method emerged relatively recently in the development of im-
age synthesis. Radiosity methods represent the development of several trends:
the development of physically based shading models, the use of more rigorous
computational methods, and the continuing tension between interactivity and re-
alism in computer graphics. The historical development of image synthesis and
radiosity will be discussed in this section.
CRTs were used as computer displays as early as the late 1940s. Such de-
vices were capable of drawing dots and lines (vectors) on the CRT according
to coordinates provided by the computer. Ivan Sutherlands Sketchpad program
[228], an interactive 2D drawing application, provided an important demonstra-
tion of the potential of interactive computer graphics. Subsequent years saw

Radiosity and Realistic Image Synthesis 4

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.2. A SHORT HISTORICAL PERSPECTIVE

many developments in vector graphics, including methods for representing and

manipulating free-form curved surfaces for applications such as mechanical and
industrial design.

1.2.1 Raster Graphics

By the late 1960s, the price of computer memory decreased to the point where
raster graphics became practical. In raster graphics the computer specifies colors
for an array of picture elements, or pixels, instead of drawing vectors, thus
allowing the more realistic portrayal of surfaces. The seminal work of Bouknight
[37], Gouraud [103], and Phong [182] explored the use of shading models to
characterize surface shape visually. The models were ad hoc, in that they were
not derived from physical models of light reflection. The models were also local,
in that they computed shading based only on the relative positions of the light,
the surface, and the eye. Illumination due to light reflected from other surfaces
was ignored, as were other global phenomena such as the shadowing of one
surface by another. In color plate 1, which contains six renderings of a simple
environment computed using various algorithms, color plate 1a is rendered using
a simple local shading model.
Another preoccupation of early researchers was the problem of determining
the visible surfaces in an image; a wide variety of algorithms were developed
for this purpose. Although visibility was originally posed as the problem of
determining what is seen by the eye, visible surface algorithms turn out to be
important to shading in general (e.g., in determining the surfaces that are visible
to a light source).
Much of this early work was directed towards improving the information
conveyed by interactive graphics. Thus, the primary objective was efficiency
of computation as opposed to accurate physical simulation. As stated by Phong
[182]:

We do not expect to be able to display the object exactly as it would

appear in reality, with texture, overcast shadows, etc. We hope only
to display an image that approximates the real object closely enough
to provide a certain degree of realism.

The success of these early local illumination models and visibility algorithms
is attested to by the presence of their direct descendants in the microcode and
hardware of current graphics workstations. Such workstations are currently
capable of displaying on the order of one million shaded polygons per second.
In spite of the focus on interactive graphics, the ultimate attraction of realism
was not lost on early researchers. Appel [8] recognized that

Radiosity and Realistic Image Synthesis 5

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.2. A SHORT HISTORICAL PERSPECTIVE

. . . many difficult problems need to be solved such as the effect

of illumination by direct and diffuse lighting, atmospheric diffusion,
back reflection, the effect of surface texture, tonal specification and
transparency of surfaces . . .

Early steps toward solving these problems were taken with the develop-
ment of techniques like texture mapping and bump mapping [31, 32, 44], which
allowed the realistic representation of more complex surface properties. In ad-
dition, visible surface algorithms were applied to the problem of determining
shadows [13, 36, 67].

1.2.2 Global Illumination Models

As Appel recognized, greater realism requires global illumination models, which
account for the interreflection of light between surfaces. It was not until 1980
that the first global illumination algorithm was introduced by Whitted [265].
Whitteds innovation was the recursive application of ray tracing to evaluate
a simple global illumination model accounting for mirror reflection, refraction,
and shadows. The resulting spectacular images inspired growing interest in
photorealism.
Whitted recognized that the evaluation of a global illumination model re-
quires determining the surfaces visible in various directions from the point to
be shaded. The heart of the ray tracing algorithm is thus the point visibility test
provided by ray casting. Much of the subsequent innovation in ray tracing has
consisted of faster algorithms for performing this visibility test.
The basic ray tracing strategy was extended to glossy reflection and soft
shadows using stochastic ray tracing [63, 64] and cone tracing [7]. Color plate
1b was rendered using stochastic ray tracing to compute illumination from the
area light source in the ceiling and glossy reflection on the floor. Although
ray traced images continued to improve, the accuracy of the simulations was
difficult to quantify since the reflection and illumination models were not based
on physical principles and quantities. Also, ray tracing did not provide a practical
strategy for computing diffuse interreflection.
More accurate physically based local reflection models were developed by
Blinn [30] and Cook and Torrance [65], using results from the fields of radiative
heat transfer and illumination engineering. This work contributed to a clearer
understanding of the appropriate physical quantities for illumination, as well
as an increased awareness of the results available in the engineering and the
physical sciences.

Radiosity and Realistic Image Synthesis 6

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.2. A SHORT HISTORICAL PERSPECTIVE

1.2.3 Early Radiosity Methods

In 1984, researchers at Fukuyama and Hiroshima Universities in Japan and at the
Program of Computer Graphics at Cornell University in the United States began
to apply radiosity methods from the field of radiative heat transfer to image
synthesis. These methods were first developed in the l950s for computing
radiant interchange between surfaces [216], for engineering applications ranging
from radiator and boiler design to the analysis of radiative transfer between
panels on spacecraft.
In image synthesis, radiosity2 methods are applicable to solving for the
interreflection of light between ideal (Lambertian) diffuse surfaces. Initial al-
gorithms [100] were restricted to environments in which all surfaces could see
each other. In following years, radiosity algorithms allowing occlusion were de-
veloped [60, 175], and efficiency was improved through the use of a hierarchical
subdivision of the environment [61, 116].
Radiosity is a departure for image synthesis for several reasons. As opposed
to the earlier empirical techniques, radiosity begins with an energy balance equa-
tion, which is then approximated and solved by numerical means. In contrast
to ray tracing, which evaluates the illumination equation for directions and lo-
cations determined by the view and the pixels of the image, radiosity solves the
illumination equation at locations distributed over the surfaces of the environ-
ment. This specification of the unknowns is independent of the viewer position,
and thus radiosity methods are often called view-independent techniques. Of
course, a final image is dependent on the viewer position and the screen reso-
lution, but most of the computational effort is complete before the selection of
viewing parameters. In this way, efficient interactive walkthroughs of simulated
environments can be performed following the radiosity preprocess. Color plate
14 shows an early radiosity solution by Nishita and Nakamae. The effect of
including indirect illumination by diffusely interreflected light is apparent when
this image is compared to the image in color plate 11, in which only direct
illumination is accounted for.
While the original radiosity method is based on the assumption of Lamber-
tian diffuse reflection, subsequent work has included extensions of the radiosity
approach to glossy and ideal (mirror) reflection [132, 217, 218, 246]. Rushmeier
[200] has also exceeded the basic radiosity formulation to include participating
media (e.g., smoke and haze). Color plates 1c-1e were rendered using varia-
tions of the radiosity method. Color plate 1c is the result of the original radiosity
method for diffuse environments. Note that indirect illumination adds color to
2
The term radiosity refers to a measure of radiant energy, in particular, the energy
leaving a surface per unit area per unit time. Over time, radiosity has also come to mean
a set of computational techniques for computing global illumination.

Radiosity and Realistic Image Synthesis 7

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.3. RADIOSITY AND FINITE ELEMENT METHODS

the shadows and the shadowed faces of the boxes. Color plate 1d is the result
of extensions that provide glossy reflection on the floor, while Color plate 1e
includes the effect of smoke within the environment.
More recent work has directly addressed the computational complexity of
radiosity algorithms. In 1988, Cohen et al. [59] introduced a progressive re-
finement approach that allows fast approximate solutions to be displayed. In
1991, Hanrahan et al. [116] formulated a complete hierarchical radiosity system
leading to a linear time algorithm. A great deal of work has also been devoted
to the critical step of discretizing or meshing the surfaces [21, 43, 154, 230]. An
important recent trend has been the incorporation of quantitative error estimates
into the solution process. Examples include estimates of integration error [19]
and the use of geometricand energy-based error metrics in the hierarchical
algorithm of Hanrahan et al. [116].

1.2.4 The Rendering Equation

Kajiya [135] unified the discussion of global illumination algorithms in 1986
with the general rendering equation. Kajiya applied Monte Carlo integration
methods to solving the rendering equation and proposed a number of techniques
for accelerating the convergence of the solution. Color plate 1f was rendered
using a Monte Carlo solution to the rendering equation.

1.3 Radiosity and Finite Element Methods

Radiosity can be understood as a particular approach to solving the rendering
equation under the assumption of Lambertian diffuse reflection. Heckbert and
Winget [125] have shown that radiosity is essentially a finite element method.
Like Monte Carlo techniques, the finite element method is a broadly ap-
plicable approach to solving difficult integral equations, such as the rendering
equation. The basic approach is to approximate an unknown function by subdi-
viding the domain of the function into smaller pieces or elements, across which
the function can be approximated using relatively simple functions like poly-
nomials. The unknown function is thus projected into a finite function space,
in which the approximated function is fully characterized by a finite number of
unknowns. The resulting system can then be solved numerically.
The ideas underlying the finite element method were first discussed as early
as the 1940s [66], although the term finite element did not become popular until
the 1960s [57]. The development of the finite element method closely paralleled
related work in approximating functions using piecewise polynomials or splines
[205]. It was also recognized in the 1950s that finite element methods were a
form of the more general Ritz variational methods.

Radiosity and Realistic Image Synthesis 8

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.3. RADIOSITY AND FINITE ELEMENT METHODS

Eye

A powerful demonstration, introduced by Goral [105], of the differences

between radiosity methods and traditional ray tracing is provided by ren-
derings of a sculpture, Construction in Wood, A Daylight Experiment, by
John Ferren (color plate 2). The sculpture, diagramed above, consists of a
series of vertical boards painted white on the faces visible to the viewer.
The back faces of the boards are painted bright colors. The sculpture is illu-
minated by light entering a window behind the sculpture, so light reaching
the viewer first reflects off the colored surfaces, then off the white surfaces
before entering the eye. As a result, the colors from the back of the boards
bleed onto the white surfaces. Color plates 2-4 show a photograph of
the sculpture and ray tracing and radiosity renderings of the sculpture. The
sculpture is solid white in the ray traced image since illumination due to
diffuse interreflection is ignored. The radiosity method, however, accounts
for the diffuse interreflections and reproduces the color bleeding.

It was not until computers became more routinely available in the 1960s and
1970s that these methods became a common technique for engineering analysis.
Since then, there has been considerable research resulting in many working
finite element codes and in a better theoretical understanding of convergence
and other mathematical properties of such methods. In addition, a number of
excellent texts have also been written [23, 70, 273].

As Heckbert and Winget [125] point out, the heat transfer formulations
upon which radiosity is based can be viewed as simple finite element methods.

Radiosity and Realistic Image Synthesis 9

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.4. THE RADIOSITY METHOD AND THIS BOOK

Heckbert and Winget emphasize the need for quantitative error metrics and show
that an explicit finite element approach considerably clarifies the understanding
of the accuracy of the approximation. Radiosity will be presented in this book
as a finite element method. However, this book cannot begin to do justice to
the broad field of finite element methods in general, and the reader is referred
to the above-mentioned texts for a wider theoretical background, as well as for
a wealth of practical information.

1.4 The Radiosity Method and This Book

This book is structured as follows (see Figure 1.2 for a diagram of the book's
structure). The first step is to derive a mathematical model of global illumination.
This derivation is undertaken in Chapter 2, working from basic transport theory
to the rendering equation, and finally making the assumptions that lead to the
radiosity equation.
In Chapter 3, the basic principles of finite element approximation are used
to cast the radiosity equation into a discrete form that is amenable to numerical
solution. In particular, the original radiosity function is approximated by a sum
of weighted basis functions. These basis functions are in turn defined by a mesh
or discretization of the surfaces in the environment.
The finite element formulation of the radiosity integral equation produces
a system of linear equations that must be solved for the weights of the basis
functions. The coefficients of this linear system are formed by integrals over
portions of the surfaces in the environment. These integrals can be solved using
both analytic and numeric methods. Chapter 4 describes a variety of algorithms
that have been developed for this purpose.
Techniques for solving the matrix equation once it has been formulated are
described in Chapter 5. We will examine a number of linear equation solvers and
discuss their applicability to the system of equations resulting from the radiosity
problem.
Chapters 6, 7 and 8 cover the general problem of subdividing the surfaces
of the model into the elements upon which the finite element approximation is
based. The accuracy and the efficiency of the solution are strongly dependent
on this subdivision. Basic subdivision strategies are described in Chapter 6. The
use of hierarchical methods that incorporate subdivision into the solution process
itself and accelerate the matrix solution is described in Chapter 7. Chapter 8
covers the basic mechanics of meshing.
Once a solution has been obtained, the final step is to produce an image,
which is discussed in Chapter 9. This is less straightforward than it might seem,
due to the limitations of display devices and the demands of visual perception.

Radiosity and Realistic Image Synthesis 10

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 1. INTRODUCTION
1.4. THE RADIOSITY METHOD AND THIS BOOK

Figure 1.2: Diagram of the radiosity method indicating the chapters where con-
cepts are discussed.

In Chapter 10 techniques for extending the basic radiosity method are de-
scribed. These provide methods to handle more general global illumination
models, including general light sources, glossy and mirror reflection, and par-
ticipating media. With these more general approaches, the distinction between
ray tracing and radiosity will become less clear.
Chapter 11 concludes this book with a discussion of applications that are
already taking advantage of this technology. We also discuss current trends in
the development of radiosity methods.
Another way to look at the organization of the book is to relate it to the
flow of information in a generic radiosity algorithm. This view is provided by
the diagram in Figure 1.2.

Radiosity and Realistic Image Synthesis 11

Edited by Michael F. Cohen and John R. Wallace
.
CHAPTER 2. RENDERING CONCEPTS

Chapter 2

Rendering Concepts
by Pat Hanrahan

2.1 Motivation
The progress in rendering in the last few years has been driven by a deeper and
better understanding of the physics of materials and lighting. Physically based or
realistic rendering can be viewed as the problem of simulating the propagation
of light in an environment. In this view of rendering, there are sources that
emit light energy into the environment; there are materials that scatter, reflect,
refract, and absorb light; and there are cameras or retinas that record the quantity
of light in different places. Given a specification of a scene consisting of the
positions of objects, lights and the camera, as well as the shapes, material, and
optical properties of objects, a rendering algorithm computes the distribution of
light energy at various points in the simulated environment.
This model of rendering naturally leads to some questions, the answers to
which form the subjects of this chapter.

1. What is light and how is it characterized and measured?

2. How is the spatial distribution of light energy described mathematically?

3. How does one characterize the reflection of light from a surface?

4. How does one formulate the conditions for the equilibrium flow of light
in an environment?

In this chapter these questions are answered from both a physical and a
mathematical point of view. Subsequent chapters will address specific represen-
tations, data structures, and algorithms for performing the required calculations
by computer.

Radiosity and Realistic Image Synthesis 13

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.2 BASIC OPTICS

Figure 2.1: Electromagnetic spectrum.

2.2 Basic Optics

Light is a form of electromagnetic radiation, a sinusoidal wave formed by cou-
pled electric and magnetic fields. The electric and magnetic fields are perpen-
dicular to each other and to the direction of propagation. The frequency of
the oscillation determines the wavelength. Electromagnetic radiation can exist
at any wavelength. From long to short, there are radio waves, microwaves,
infrared, light, ultraviolet, x-rays, and gamma rays (see Figure 2.1).
A pure source of light, such as that produced by a laser, consists of light
at a single frequency. In the natural world, however, light almost always exists
as a mixture of different wavelengths. Laser light is also coherent, that is, the
source is tuned so that the wave stays in phase as it propagates. Natural light,
in contrast, is incoherent.
Electromagnetic radiation can also be polarized. This refers to the preferen-
tial orientation of the electric and magnetic field vectors relative to the direction
of propagation. Just as incoherent light consists of many waves that are summed
with random phase, unpolarized light consists of many waves that are summed
with random orientation. The polarization of the incident radiation is an impor-
tant parameter affecting the reflection of light from a surface, but the discussion
will be simplified by ignoring polarization.
The fact that light is just one form of electromagnetic radiation is of great
benefit for computer graphics in that it points to theory and algorithms from
many other disciplines, in particular, optics, but also more applied disciplines
such as radar engineering and radiative heat transfer. The study of optics is typ-
ically divided into three subareas: geometrical or ray optics, physical or wave
optics, and quantum or photon optics. Geometrical optics is most relevant to
computer graphics since it focuses on calculating macroscopic properties of light

Radiosity and Realistic Image Synthesis 14

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.3 RADIOMETRY AND PHOTOMETRY

as it propagates through environments. Geometrical optics is useful to under-

stand shadows, basic optical laws such as the laws of reflection and refraction,
and the design of classical optical systems such as binoculars and eyeglasses.
However, geometrical optics is not a complete theory of light. Physical or wave
optics is necessary to understand the interaction of light with objects that have
sizes comparable to the wavelength of the light. Physical optics allows us to
understand the physics behind interference, dispersion, and technologies such as
holograms. Finally, to explain in full detail the interaction of light with atoms
and molecules quantum mechanics must be used. In the quantum mechanical
model light is assumed to consist of particles, or photons. For the purposes of
this book, geometrical optics will provide a full-enough view of the phenomena
simulated with the radiosity methods.

2.3 Radiometry and Photometry

Radiometry is the science of the physical measurement of electromagnetic en-
ergy. Since all forms of energy in principle can be interconverted, a radiometric
measurement is expressed in the SI units for energy or power, joules and watts,
respectively. The amount of light at each wavelength can be measured with a
spectroradiometer, and the resulting plot of the measurements is the spectrum
of the source.
Photometry, on the other hand, is the psychophysical measurement of the
visual sensation produced by the electromagnetic spectrum. Our eyes are only
sensitive to the electromagnetic spectrum between the ultraviolet (380 nm) and
the infrared (770 nm). The most prominent difference between two sources of
light with different mixtures of wavelengths is that they appear to have different
colors. However, an equally important feature is that different mixtures of light
also can have different luminosities, or brightnesses.
Pierre Bouguer established the field of photometry in 1760 by asking a hu-
man observer to compare different light sources [35]. By comparing an unknown
source with a standard source of known brightnessa candle at the timethe
relative brightness of the two sources could be assessed. Bouguers experiment
was quite ingenious. He realized that a human observer could not provide an
accurate quantitative description of how much brighter one source was than an-
other, but could reliably tell whether two sources were equally bright.1 Bouguer
was also aware of the inverse square law. Just as Kepler and Newton had used it
to describe the gravitational force from a point mass source, Bouguer reasoned
that it also applied to a point light source. The experiment consisted of the
1
This fact will be used in Chapter 9 when algorithms to select pixel values for display
are examined.

Radiosity and Realistic Image Synthesis 15

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.3 RADIOMETRY AND PHOTOMETRY

Figure 2.2: Spectral luminous relative efficiency curve.

observer moving the standard source until the brightnesses of the two sources
were equal. By recording the relative distances of the two light sources from
the eye, the relative brightnesses can be determined with the inverse square law.
Bouguer founded the field of photometry well before the mechanisms of
human vision were understood. It is now known that different spectra have
different brightnesses because the pigments in our photoreceptors have different
sensitivities or responses toward different wavelengths. A plot of the relative
sensitivity of the eye across the visible spectrum is shown in Figure 2.2; this
curve is called the spectral luminous relative efficiency curve. The observers
response, R, to a spectrum is then the sum, or integral, of the response to each
spectral band. This in turn is equal to the amount of energy at that wavelength,
l, times its relative luminosity.

770
R= V( )S( )d (2.1)
380 nm

where V is the relative efficiency and S is the spectral energy. Because there
is wide variation between peoples responses to different light sources, V has
been standardized.
Radiometry is more fundamental than photometry, in that photometric quan-
tities may be computed from spectroradiometric measurements. For this reason,
it is best to use radiometric quantities for computer graphics and image syn-
thesis. However, photometry preceded radiometry by over a hundred years, so
much of radiometry is merely a modern interpretation of ideas from photometry.
As mentioned, the radiometric units for power and energy are the watt and
joule, respectively. The photometric unit for luminous power is the lumen,
and the photometric unit for luminous energy is the talbot. Our eye is most

Radiosity and Realistic Image Synthesis 16

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4 THE LIGHT FIELD

sensitive to yellow-green light with a wavelength of approximately 555 nm that

has a luminosity of 684 lumens per watt. Light of any other wavelength, and
therefore any mixture of light, will yield fewer lumens per watt. The number of
lumens per watt is a rough measure of the effective brightness of a light source.
For example, the garden-variety 40-Watt incandescent light bulb is rated at only
490 lumensroughly 12 lumens per watt. Of course, the wattage in this case
is not the energy of the light produced, but rather the electrical energy consumed
by the light bulb. It is not possible to convert electrical energy to radiant energy
with 100% efficiency so some energy is lost to heat.
When we talk about light, power and energy usually may be used inter-
changeably, because the speed of light is so fast that it immediately attains
equilibrium. Imagine turning on a light switch. The environment immediately
switches from a steady state involving no light to a state in which it is bathed
in light. There are situations, however, where energy must be used instead of
power. For example, the response of a piece of film is proportional to the total
energy received. The integral over time of power is called the exposure. The
concept of exposure is familiar to anyone who has stayed in the sun too long
and gotten a sunburn.
An important principle that must be obeyed by any physical system is the
conservation of energy. This applies at two levelsa macro or global level, and
a micro or local level.

At the global level, the total power put into the system by the light sources
must equal the power being absorbed by the surfaces. In this situation
energy is being conserved. However, electrical energy is continuing to
flow into the system to power the lights, and heat energy is flowing out
of the system because the surfaces are heated.

At the local level, the energy flowing into a region of space or onto a
surface element must equal the energy flowing out. Accounting for all
changes in the flow of light locally requires that energy is conserved.
Thus, the amount of absorbed, reflected, and transmitted light must never
be greater than the amount of incident light. The distribution of light can
also become more concentrated or focused as it propagates. This leads to
the next topic which is how to characterize the flow of light.

2.4 The Light Field

2.4.1 Transport Theory

The propagation of light in an environment is built around a core of basic ideas
concerning the geometry of flows. In physics the study of how stuff flows

Radiosity and Realistic Image Synthesis 17

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4 THE LIGHT FIELD

Figure 2.3: Particles in a differential volume.

is termed transport theory. The stuff can be mass, charge, energy, or light.
Flow quantities are differential quantities that can be difficult to appreciate and
manipulate comfortably. In this section all the important physical quantities
associated with the flow of light in the environment will be introduced along
with their application to computer graphics.
The easiest way to learn transport quantities is to think in terms of particles
(think of photons). Particles are easy to visualize, easy to count, and therefore
easy to track as they flow around the environment. The particle density p(x) is
the number of particles per unit volume at the point x (see Figure 2.3). Then
the total number of particles, P(x), in a small differential volume dV is

P(x) = p(x) dV (2.2)

Note that the particle density is an intrinsic or differential quantity, whereas the
total number of particles is an absolute or extrinsic quantity.
r Now imagine a stream of particles all moving with the same velocity vector
v ; that is, if they are photons, not only are they all moving at the speed of
light, but they are all moving in the same direction. We wish to count the total
number of particles flowing across a small differential surface element dA in
a slice of time dt. The surface element is purely fictitious and introduced for
r
convenience and may or may not correspond to a real physical surface. In time
dt each particle moves a distance v dt. How many particles cross dA? This
can be computed using the following observation: r consider the tube formed by
v
sweeping dA a distance vdt in the direction . All particles that cross dA
between t and t + dt must have initially been inside this tube at time t. If they
were outside this tube, they would not be moving fast enough to make it to
the surface element dA in the allotted time. This implies that one can compute
the number of particles crossing the surface element by multiplying the particle
volume density times the volume of the tube. The volume of the tube is just
equal to its base (dA) times its height, which is equal to v cos dt. Therefore,
as depicted in Figure 2.4, the total number of particles crossing the surface is

P(x) = p(x) dV
= p(x)(v dt cos ) dA (2.3)
Radiosity and Realistic Image Synthesis 18
Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4. THE LIGHT FIELD

Figure 2.4: Total particles crossing a surface.

Note that the number of particles flowing through a surface element depends
on both the area of the surface element and its orientation relative to the flow.
Observe that the maximum flow through a surface of a fixed size occurs when
the surface is oriented perpendicular to the direction of flow. Conversely, no
particles flow across a surface when it is oriented parallel to the flow. More
specifically, the above formula says that the flow across a surface depends on the
cosine of the angle of incidence between the surface normal and the direction of
the flow. This fact follows strictly from the geometry of the situation and does
not depend on what is flowing.
The number of particles flowing is proportional both to the differential area
of the surface element and to the interval of time used to tally the particle count.
If either the area or the time interval is zero, the number of particles flowing
is also zero and not of much interest. However, we can divide through by the
time interval dt and the surface area dA and take the limit as these quantities
go to zero. This quantity is called the flux.
More generally all the particles through a point will not be flowing with
the same speed and in the same direction. Fortunately, the above calculation is
fairly easy to generalize to account for a distribution of particles with different
velocities moving in different directions. The particlerdensity is now a function
of two independent variables, position x and direction . Then, just as before, the
r
number of particles flowing across a differential surface element in the direction
equals r r
P(x, ) = p(x, ) cos d dA (2.4)

Here the notation d is introduced for the differential solid angle. The direction
of this vector is in the direction of the flow, rand its length is equal to the small
differential solid angle of directions about . For those unfamiliar with solid
angles and differential solid angles, please refer to the box.

2.4.2 Radiance and Luminance

The above theory can be immediately applied to light transport by considering
ight as photons. However, rendering systems almost never need consider (or at

Radiosity and Realistic Image Synthesis 19

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4 THE LIGHT FIELD

Angles and Solid Angles

A direction is indicated by the vector . Since this is a unit vector, it can be

represented by a point on the unit sphere. Positions on a sphere in turn can
be represented by two angles: the number of degrees from the North Pole
or zenith, , and the number of degrees about the equator or azimuth, .
r
Directions and spherical coordinates ( , ) can be used interchangeably.

A big advantage of thinking of directions as points on a sphere comes when

considering differential distributions of directions. A differential distribution
of directions can be represented by a small region on the unit sphere.

least have not considered up to this point) the quantum nature of light. Instead,
when discussing light transport, the stuff that flows, or flux, is the radiant energy
per unit time, or radiant power , rather than the number of particles. The radiant
energy per unit volume is simply the photon volume density times the energy of
a single photon h c/, where h is Plancks constant and c is the speed of light.
The radiometric term for this quantity is radiance.
r r hc
L(x, ) = p(x, , ) d (2.6)

Radiance is arguably the most important quantity in image synthesis. Defined
precisely, radiance is power per unit projected area perpendicular to the ray per
unit solid angle in the direction of the ray (see Figure 2.5). The definition in
equation 2.6 is that proposed by Nicodemus [174], who was one of the first
authors to recognize its fundamental nature.
The radiance distribution completely characterizes the distribution of light

Radiosity and Realistic Image Synthesis 20

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4. THE LIGHT FIELD

The area of a small differential surface element on a sphere of radius r is

dA = (r d) (r sin d) = r2 sin d d

Here rd is the length of the longitudinal are generated as goes to + d.

Similarly r sin d is the length of the latitudinal are generated as goes
to + d. The product of these two lengths is the differential area of that
patch on the sphere.

This derivation uses the definition of angle in radians: the angle subtended
by a circular arc of length l is equal to l/r. The circle itself subtends an
angle of 2 radians because the circumference of the circle is 2r. By using
a similar idea we can define a solid angle. The solid angle subtended by a
spherical area a is equal to a/r2. This quantity is the measure of the angle
in steradians (radians squared), denoted sr. A sphere has a total area of
4r2, so there are 4 steradians in a sphere.
A differential solid angle, indicated as d, is then
dA
d = = sin d d (2.5)
r2
It ris very convenient torthink of the differential solid angle as a vector,
d . The direction
r of d is in the direction of the point on the sphere, and
the length of d is equal to the size of the differential solid angle in that
direction.

Figure 2.5: The radiance is the power per unit projected area perpendicular to
the ray per unit solid angle in the direction of the ray.

Radiosity and Realistic Image Synthesis 21

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4 THE LIGHT FIELD

Figure 2.6: Equality of flux leaving the first surface and flux arriving on the
second surface.

in a scene. Note that it is a function of five independent variables, three that

specify position and two that specify direction. All other radiometric quantities
can be computed from it. For example, the differential flux in a small beam
with cross-sectional area dA and differential solid angle d is
r
d = L(x, ) cos d dA (2.7)
This follows immediately from the earlier discussion of particle transport.
To emphasize further the importance of radiance, consider the following two
properties:

l. The radiance in the direction of a light ray remains constant as it prop-

agates along the ray (assuming there are no losses due to absorption or
scattering). This law follows from the conservation of energy within a
thin pencil of light, as shown in Figure 2.6.

The total flux leaving the first surface must equal the flux arriving on the
second surface.
L1 d1 dA1 = L2 d2 dA2 (2.8)
but d1 = dA2/r2 and d2 = dA1/r2, thus,
dA1 dA2
T = d1 dA1 = d2 dA2 = (2.9)
r2

Radiosity and Realistic Image Synthesis 22

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4. THE LIGHT FIELD

Figure 2.7: A simple exposure meter.

This quantity T is called the throughput of the beam; the larger the
throughput, the bigger the beam. This immediately leads to the conclusion
that
L1 = L2 (2.10)

and hence, the invariance of radiance along the direction of propagation.

As a consequence of this law, radiance is the numeric quantity that should
be associated with a ray in a ray tracer.

2. The response of a sensor is proportional to the radiance of the surface

visible to the sensor.

To prove this law, consider the simple exposure meter in Figure 2.7. This
meter has a small sensor with area a and an aperture with area A. The
response of the sensor is proportional to the total integrated flux falling
on it.
R= A L cos d dA = LT (2.11)
Thus, assuming the radiance is constant in the field of view, the response
is proportional to the radiance. The constant of proportionality is the
throughput, which is only a function of the geometry of the sensor. The
fact that the radiance at the sensor is the same as the radiance at the surface
follows from the invariance of radiance along a ray.

This law has a fairly intuitive explanation. Each sensor element sees that
part of the environment inside the beam defined by the aperture and the
receptive area of the sensor. If a surface is far away from the sensor, the
sensor sees more of it. Paradoxically, one might conclude that the surface
appears brighter because more energy arrives on the sensor. However, the
sensor is also far from the surface, which means that the sensor subtends a
smaller angle with respect to the surface. The increase in energy resulting
from integrating over a larger surface area is exactly counterbalanced by
the decrease in percentage of light that makes it to the sensor. This
property of radiance explains why a large uniformly illuminated painted
wall appears equally bright over a wide range of viewing distances.

Radiosity and Realistic Image Synthesis 23

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4 THE LIGHT FIELD

Figure 2.8: Projection of differential area.

As a consequence, the radiance from a surface to the eye is the quantity

that should be output to the display device.

2.4.3 Irradiance and Illuminance

The two properties of radiance described in the previous section were derived
by considering the total flux within a small beam of radiation. Another very
important quantity is the total energy per unit area incident onto a surface with a
fixed orientation. This can be computed by integrating the incident, or incoming
radiance, Li , over a hemisphere, .

L cos d d
d =

i (2.12)

The irradiance, E, is the radiant energy per unit area falling on a surface
(the corresponding photometric quantity is the illuminance).
d
(2.13)
d
or
=
L cos d

(2.14)

The quantity cos dw is often referred to as the projected solid angle. It can be
thought of as the projection of a differential area on a sphere onto the base of
the sphere, as shown in Figure 2.8.
This geometric construction shows that the integral of the projected solid
angle over the hemisphere is just , the area of the base of a hemisphere with

Radiosity and Realistic Image Synthesis 24

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4. THE LIGHT FIELD

unit radius. This result can also be derived directly by computing the following
integral:
2
cos d = 0 0 cos sin d d
2
= 0 0 cos d cos d
/2
cos 2
= 2
2
0
= (2.15)

Note that if all rays of light are parallel, which occurs if a single distant
source irradiates a surface, then the integral reduces to the simple formula

E = E0 cos (2.16)

where E0 is the energy per unit perpendicular area arriving from the distant
source.

2.4.4 Radiosity and Luminosity

As the title of this book suggests, radiosity is another important quantity in
image synthesis. Radiosity, B, is very similar to irradiance. Whereas irradiance
is the energy per unit area incident onto a surface, radiosity is the energy per
unit area that leaves a surface. It equals

B= Lo cos d (2.17)

where Lo is the outgoing radiance.

The official term for radiosity is radiant exitance. Because of the wide-
spread use of the term radiosity in the computer graphics literature, it will be
used in this book. The photometric equivalent is luminosity.

2.4.5 Radiant and Luminous Intensity

Radiance is a very useful way of characterizing light transport between surface
elements. Unfortunately, it is difficult to describe the energy distribution of a
point light source with radiance because of the point singularity at the source.
Fortunately, it is very easy to characterize the energy distribution by introducing
another quantitythe radiant or luminous intensity.

Radiosity and Realistic Image Synthesis 25

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4 THE LIGHT FIELD

Note that this use of intensity is very different from that typically used
by the computer graphics community. Even more confusion results because
intensity is often used to indicate radiance-like transport quantities in the physics
community. The radiant intensity is quite similar to that used in the geometric
optics community.
The energy distribution from a point light source expands outward from the
center. A small beam is defined by a differential solid angle in a given direction.
The flux in a small beam dw is defined to be equal to
r
d I ( ) d (2.18)
I is the radiant intensity of the point light source with units of power per unit
solid angle. The equivalent photometric quantity is the luminous intensity.
The radiant intensity in a given direction is equal to the irradiance at a
point on the unit sphere centered at the source. In the geometric optics literature
intensity is defined to be the power per unit area (rather than per unit solid angle).
In the case of a spherical wavefront emanating from a point source, the geometric
optics definition is basically the same as the radiometric definition. However,
in general, the wavefront emanating from a point source will be distorted after
it reflects or refracts from other surfaces and so the definition in terms of solid
angle, is less general.
For an isotropic point light source,

I = (2.19)
4
Of course, a point source may act like a spotlight and radiate different amounts
of light in different directions. The total energy emitted is then
r
= I ( ) d (2.20)
The irradiance on a differential surface due to a single point light source can
be computed by calculating the solid angle subtended by the surface element
from the point of view of the light source.
d cos
E = I = (2.21)
d 4 x x 2
S

where x x S is the distance from the point to the surface element. Note the
1/r2 fall-off: this is the origin of the inverse square law.
The distribution of irradiance on a surface is often drawn using a contour
plot or iso-lux diagram, while the directional distribution of the intensity from
a point light source is expressed with a goniometric or iso-candela diagram.2
This is a contour plot of equal candela levels as a function of the (,).
2
See Chapter 10 for details of lighting specifications.

Radiosity and Realistic Image Synthesis 26

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.4. THE LIGHT FIELD

Physics Radiometry Radiometric Units

Radiant energy joules [J = kgm2/s2]

Flux Radient power watss [W = joules/s]
Angular flux density Radiance [W/m2sr]
Flux density Irradiance [W/m2]
Flux density Radiosity [W/m2]
Radiant intensity [W/sr]

Physics Photometry Photometric Units

Luminous energy talbot

Flux Luminous power lumens [talbot/second]
Angular flux density Luminance Nit [lumens/m2sr]
Flux density Illuminance Lux [lumens/m2sr]
Flux density Luminosity Lux [lumens/m2sr]
Luminous intensity Candela [lumens/sr]

Table 2.1: Radiometric and photometric quantities.

2.4.6 Summary of Radiometric and Photometric Quantities

In most computer graphics systems, optical quantities are simply colors denoted
by red, green, and blue triplets. These triplets are used to specify many quanti-
ties including light sources, material properties, and intermediate calculations.3
As noted, there is a small but finite number (six to be exact) of radiometric
(photometric) quantities that characterize the distribution of light in the environ-
ment. They are the radiant energy (luminous energy), radiant power (luminous
power), radiance (luminance), irradiance (illuminance), radiosity (luminosity),
and radiant intensity (luminous intensity). These quantities and their units are
summarized in Table 2.1.

3
A more complete treatment of color specification is given in Chapter 9.

Radiosity and Realistic Image Synthesis 27

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

Figure 2.9: Bidirectional reflection distribution function.

2.5 Reflection Functions

The next question is how to characterize the reflection of light from a surface.
Reflection is defined as the process by which light incident on a surface leaves
that surface from the same side. Transmission, absorption, spectral and polariza-
tion effects, fluorescence, and phosphorescence are also important to consider in
developing an accurate model of the interaction of light with materials, but will
not be treated in detail here. Instead, this section will concentrate on nomencla-
ture and the general properties that are satisfied by all reflection functions.

2.5.1 The Bidirectional Reflection Distribution Function

Consider the lightr incident on a surface from a small differential solid angle
r
in the direction i. The amount of reflectedr light in another direction r is
proportional to the incident irradiance from i (see Figure 2.9). That is,
r r
dLr ( r ) dE ( i ) (2.22)

Equation 2.22 simply states that an increase in the incident light energy per unit
area results in a corresponding increase in the reflected light energy. The incident
irradiance can be increased by increasing either the solid angle subtended by the
source or the energy density in the beam.
The constant of proportionality is termed the bidirectional reflection distri-
bution function, or BRDF.
r
r r Lr ( r )
f r ( i r ) r (2.23)
Li ( i ) cos i d i
Radiosity and Realistic Image Synthesis 28
Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

Figure 2.10: Helmholtz reciprocity principle.

More precisely,r the BRDF is defined to be the ratio of the reflected radiance in
the direction r to the differential irradiance from the incident direction i that
produces it. The BRDF is bidirectional because it depends on two directions.
Often, the dependence on the four angles is made explicit by writing the BRDF
as fr(i , i ; r , r ). The BRDF is a distribution function because it is strictly
positive. Since it gives the concentration of flux per steradian, it may take on
any value between zero and infinity. The BRDF has units of inverse steradians.
The BRDF has several interesting properties:

1. If the BRDF is based on physical laws, then it will remain unchanged if

the incident and reflected directions are interchanged. That is,
r r r r
fr ( r i ) = fr ( i r ) (2.24)
This Helmholtz reciprocity principle is equivalent to saying that if a photon
moves along a path, it will follow the same path if its direction is reversed
(see Figure 2.10).

2. The BRDF is, in general, anisotropic. That is, if the incident and reflected
directions are fixed and the underlying surface is rotated about the surface
normal, the percentage of light reflected may change (see Figure 2.11).
Examples of anisotropic materials are brushed aluminum or cloth [134].

Many materials, however, are smooth and their reflectivity does not depend
on the surfaces orientation. Thus, their reflection functions do not change
if the surface is rotated, and
fr ((i , i + ) (r , r + )) = fr ((i , i) (r , r)) (2.25)
This implies that the reflection function has only three degrees of freedom
instead of four.

Radiosity and Realistic Image Synthesis 29

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

Figure 2.11: The reflection may change with rotations of the surface due to
anisotropy.

Notice that adding light from another incident direction has no influence
on the amount of light reflected from other incident directions. Thus, reflection
behaves linearly, and hence the total amount of light reflected by a surface in a
specific direction is given by a hemispherical integral over all possible incident
directions. This leads to the reflectance equation:
r r r r
Lr ( r ) = i
f r ( i r ) Li ( i ) cos i d i (2.26)

Put another way, the reflected radiance in a particular direction is due to the
radiance arriving from all directions weighted by the BRDF relating the incoming,
and reflected directions and by the projected solid angle.

2.5.2 Mirror Reflection

As an example of a BRDF, consider a perfect mirror and the geometry of the
reflection. For a mirror, the angle of reflectance is equal to the angle of incidence,
and the reflected vector is in the plane determined by the incident ray and surface
normal vector. This implies that

r = i
r = i (2.27)

Second, consider the radiometry of reflection. For a mirror, the reflected

radiance is exactly equal to the incident radiance.

Lr (r , r ) = Li (r , r ) (2.28)

This physical fact can be mathematically expressed with a BRDF involving delta

Radiosity and Realistic Image Synthesis 30

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

functions.
(cos i cos r )
f r ,m = ( i ( r )) (2.29)
cos i
Recall that the delta function has the following three properties:
1. (x) = 0 if x 0

2.
(x) dx = 1

3.

(x y) f (x)dx = f ( y)
It can be verified that this leads to the correct reflected radiance by performing
the hemispherical integral.
(cos i cos r )
Lr ( r , r ) = ( i ( r ))
i cos i
Li ( i , i ) cos i d i d i (2.30)
= L (r , r )
i

2.5.3 The Reflectance

Recall that the delta function can be interpreted as an infinitesimally thin, in-
finitely high spike with unit area. This implies that the BRDF, although always
positive, may be infinite. Often it is more intuitive to work with a quantity that
is bounded between 0 and 1. This quantity is called the biconical reflectance,
or simply reflectance.
Consider the ratio of reflected flux to incident flux. Since the reflected flux
must always be less than the incident flux giving rise to it, the reflectance must
always be less than 1.
r
d r r Lr ( r ) cos r d r
= r
d i i Li ( i ) cos i d i
r r r
r i f r ( i r ) Li ( i ) cos i d i cos r d r (2.31)
= r
Li ( i ) cos i d i
i

Unfortunately, the reflectance depends on the distribution of incoming light, Li .

If it is assumed that Li is uniform and isotropic, then Li can be taken out from
the integral in both the numerator and the denominator. This results in the
relationship between the reflectance and the BRDF which forms the definition
of the reflectance:
r r
r r r i f r ( i r ) cos i d i cos r d r
r ( i r ) (2.32)
cos i d i i

Radiosity and Realistic Image Synthesis 31

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

The reflectance involves a double integral over the incident and reflected
directions for which the limits of integration have not yet been set. Three
choices for the limits are a differential solid angle, a finite solid angle, or the
entire hemisphere. Since this choice can be made for both the incident and the
reflected directions, there are nine different reflectances. These are shown in
table 2.2.

2

(i r) ( r )
i ( 2 )
i

(i r) (i r) (i 2)

2 (2 r) (2 r) (2 2)

Table 2.2: The nine biconical reflectances.

The names of these reflectances are formed by combining the following

words: directional (for differential solid angle), conical (for finite solid angle),
andr hemispherical
r (for a solid angle equal to the entire hemisphere). Thus,
( i r ), (i r ), and (2 2) are referred to as the bidi-
rectional, biconical, and bihemispherical reflectances, respectively. Perhaps the
mostr interesting reflectance function is the directional-hemispherical reflectance,
( i, 2). This is the amount of light scattered into the entire hemisphere
from a single incident direction. Since this quantity is the ratio of fluxes, it must
be less, than 1. However, be aware that this quantity can change with the angle
of incidence.

2.5.4 Lambertian Diffuse Reflection

To illustrate the relationship between the BRDF and reflectance, consider the case
of Lambertian diffuse reflectance. Diffuse reflectance is modeled by assuming
that light is equally likely to be scattered in any direction, regardless of the
incident direction. In other words, the BRDF is constant. Thus,
r r
Lr,d ( r ) = i
f r,d Li ( i ) cos i d i
r
= f r,d Li ( i ) cos i d i
i
= f r,d E (2.33)

This leads to two conclusions:

1. The value of the reflected radiance is proportional to the incident irradi-

ance.

Radiosity and Realistic Image Synthesis 32

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

r hence the same in all directions,

2. The reflected radiance is a constant and
since neither fr,d nor E depends on r . This is true independent of the
distribution of incoming light.

The fact that energy is conserved can be ensured by forcing the hemispherical-
hemispherical reflectance to be less than 1.
r
r,d Lr,d ( r ) cos r d r
d (2 2 ) = = r
r
i i Li ( i ) cos i d i
Lr,d cos r d r
= r

E
Lr,d
=
E
= f r,d (2.34)

It thus immediately follows that if the BRDF is a constant, then the reflectance is
also a constant. More importantly, this relationship can be used to parameterize
the BRDF in terms of the reflectance: fr,d = d / . Often, it is more intuitive
to describe materials using their reflectances because they are constrained to lie
between 0 and 1. Whenever a is used in this text, it can safely be assumed to
lie between 0 and 1.
Since the outgoing radiance is constant, the radiosity
B = Lr,d (2.35)
is related to the irradiance by the following equation:
B
d = (2.36)
E
Equation 2.36 states that for diffuse reflection, the reflectance is equal to the
radiosity divided by the irradiance.

2.5.5 Glossy Reflection

In practice it is often convenient to treat the general BRDF as the sum of
three qualitatively different components: mirror (or ideal) specular reflection,
Lambertian (or ideal) diffuse reflection, and glossy reflection (see Figure 2.12).
The diffuse and mirror reflection laws, Lamberts law, and the law of reflection,
were discussed in the previous sections.
However, real materials are not perfectly diffuse or perfect mirror specular.
This is to be expected since these models of reflection are the simplest math-
ematical abstractions of the properties of surfaces and materials. Real surfaces

Radiosity and Realistic Image Synthesis 33

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

Figure 2.12: Reflectance components.

Figure 2.13: Complex reflection distributions arise from rough surface and sub-
surface phenomena.

are not planar and perfectly smooth and thus would not be expected to reflect
light in just one direction. A real BRDF will thus contain a component between
these limiting cases in which light hitting the surface from a certain direction is
reflected into a complex distribution of outgoing directions.
The terminology for the various components is highly variable in the image
synthesis literature. In particular, the intermediate component that we call glossy
reflection is variously called specular, rough specular, wide and narrow diffuse,
and directional diffuse. The term glossy has also been used in the surface reflec-
tion literature and has been selected instead for this work because its common
usage is suggestive of the intended technical meaning.
Lord Rayleigh was the first to explain the effects of surface finish on the

Radiosity and Realistic Image Synthesis 34

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.5. REFLECTION FUNCTIONS

Figure 2.14: Vectors for glossy reflection models.

reflective properties of materials. He reasoned that a surface would become

shinier if it were perfectly flat, or at least flat relative to the wavelength of the
incident radiation. His theory is relatively easy to test because the wavelengths
of common sources of radiation range from the macroscopic to the microscopic,
and it can be verified that long wavelength radiation is more easily reflected from
a surface. As shorter and shorter wavelengths are directed toward the surface,
the ideal specular component decreases and the reflection becomes less focused.
This transition occurs roughly when the wavelength of light becomes equal to
the relative height changes in the surface. Thus, glossy reflection arises from
the scattering of light from rough surfaces, an idea first proposed by Bouguer.
The mirror specular term is considered to arise from perfectly smooth surfaces,
while the Lambertian diffuse term arises from multiple surface reflections from
very rough surfaces and from subsurface scattering (see Figure 2.13).
Another important optical effect is that glossy reflection increases at glancing
angles of incidences and reflection. This is predicted by the Fresnel formula,
which gives the relative percentage of light that is reflected or refracted across a
planar boundary as a function of the angles of incidence and index of refraction.
In computer graphics glossy reflection from rough surfaces is typically mod-
eled using the microfacet theory. This theory assumes the surface is made of
little reflective facets, each behaving as a small mirror; that is, each reflecting
light perfectly. This model predicts that the amount of light reflected from a light
source toward the eye is equal to the relative number of microfacets oriented
halfway between the eye and the light source. This model has been enhanced
by many researchers [30, 65] and in its modern form consists of several terms

DGF
fr = (2.37)
4 cos r cos i

Radiosity and Realistic Image Synthesis 35

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.6. THE RENDERING EQUATION

D is the microfacet distribution. This distribution function gives the num-

ber of microfacets oriented in a particular direction. This function is
typically modeled with the following formula (see Figure 2.14):

D(H, N, ) = (N H) (2.38)

Note that this distribution is maximal when H equals N, implying that

the maximum number of microfacets are oriented parallel to the surface.
Note also that K controls the rate at which the distribution of microfacets
falls off, and is related to the roughness of the surface.

G is a geometric attenuation term accounting for self-shadowing. This

arises because a rough surface is actually a height field, and facets in the
valleys are less visible at glancing angles as facets at the peaks. This is
an important effect, but very difficult to model precisely with a simple
formula.

F is the Fresnel reflection term related to a materials index of refraction.

The modeling of reflection of light from real materials is an interesting and

important subject; however, space does not permit us to cover it in detail in
this book. Models that can be found in the literature range from Phongs simple
empirical model [181], to models of the form given above [30, 65, 236] that differ
primarily in the details of the D function, to more recent (and complex) models
such as that proposed by He et al. [118]. A good summary and description of
the earlier models is given by Hall[114]. Subsurface reflection (see Figure 2.13)
that has typically been modeled as part of the Lambertian diffuse component has
also been reexamined to provide a more physically based model for biological
materials such as skin and leaves [115].

2.6 The Rendering Equation

The reflectance equation makes it possible to compute the reflected light distri-
bution from the incident light distribution and the BRDF of the material. The
important remaining task is to specify, or preferably to compute, the incident
light distribution. This is typically referred to as the illumination model.
The first and easiest case to consider is one with no occlusion and direct
illumination from simple light sources. In this case there is typically a small
number of point or distant lights, and it can be assumed that all light arrives at
the surface; that is, there is no shadowing. Since this model does not consider
the environment as a whole and only depends on the individual properties of the
light sources and the surface being shaded, it is often called a local illumination

Radiosity and Realistic Image Synthesis 36

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.6. THE RENDERING EQUATION

model. Shadows can be added by testing whether a point on the surface is visible
to the light source. This is what is done in a ray tracer, but it requires access
to the entire environment and is therefore an example of a global illumination
model.
The second and considerably more difficult case is indirect illumination. In
this case light may come from any surface in the environment, and it is very
important to consider shadowing.
In the following sections, the interreflection of light between surfaces will
be taken into account, and the rendering equation is derived from the reflectance
equation. The radiosity equation, a simplified form of the rendering equation,
that results by assuming all surfaces are Lambertian reflectors is also derived.

2.6.1 Local or Direct Illumination

It is easy to incorporate direct lighting from point light sources into the previous
reflection models. Recall the reflectance equation
r r r r
Lr ( r ) = i
f r ( i r ) Li ( i ) cos i d i (2 .39)

Recall also that the irradiance from a single point light source was derived,
cos
E = (2.40)
4 x x 2
s
r
If the direction to the light source is given by s, then the radiance from a point
light source can be expressed with a delta function.
r
Li ( i ) = 2 (cos i cos s ) ( i s ) (2.41)
4 x xS
Substituting equation 2.41 into the reflectance equation yields
r r
r r
Lr ( r ) = f r ( i
r ) Li ( i ) cos i d i
r r (2.42)
= 2 f r ( r , s ) cos s
4 x xS
If there are n light sources, then the hemispherical integral collapses to a sum
over the n sources. This is the lighting model used by 3D graphics workstations.
It is easy to extend this model to light sources with arbitrary directional
distributions, as well as distant light sources. The above formulae are changed
to use the radiant intensity in the direction of the surface. In principle, linear
and area light sources can also be used, although this involves integrating the
reflectance function over the range of possible directions incident from the light
source. Nishita and Nakamae [175] and Amanatides [7] discuss this possibility.

Radiosity and Realistic Image Synthesis 37

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.6. THE RENDERING EQUATION

Figure 2.15: Two point transport geometry.

2.6.2 Global or Indirect Illumination

The first step in using a global illumination model is to relate the illumination on
one surface to the reflected light distribution from another surface. This requires
that the spatial dependence of radiance is made explicit and that occlusion is
considered.
Using the fact that radiance is invariant along a ray, the incident radiance
at x due to the radiance from x is
r r
Li ( x , r ) = Lo ( x, o ) V ( x, x ) (2.43)
r
where i a direction vector from x to x, and o is in the opposite direction.

r r x x
i = o = (2.44)
x x

The function V(x, x) is a visibility function. It is 1, if x and x are mutually

visible; otherwise it is 0.
Returning to the reflectance equation, the next step is to switch the hemi-
spherical integral over all incident directions to an area integral over all the
other surfaces in the environment. This is easily done by relating the solid angle
subtended by the source to its projected surface area.

cos o dA
d i = 2 (2.45)
x x

Dotting this to form the projected solid angle results in

d i cos o dA = G(x, x ) dA (2.46)

Radiosity and Realistic Image Synthesis 38

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.6. THE RENDERING EQUATION

where
cos i cos o
G(x, x ) = G( x , x) = (2.47)
x x
2

Substituting G(x, x) into the reflectance equation leads to the following

integral equation over the surfaces, S:
r
L(x, ) = S f (x) L(x, ) G(x, x) V(x, x) dA
r (2.48)

Since this equation only involves outgoing radiances and directions, the sub-
scripts denoting incoming and outgoing directions can safely be dropped (except
from fr ).
Equation 2.48 was first introduced to the computer graphics literature by
Kajiya [135], who appropriately named it the rendering equation. Actually, his
notation (and development) is slightly different than that used in equation 2.48.
He introduced a new intensity quantity, I(x x)the two point transport
intensity from x to x (see Figure 2.15). This intensity quantity is a function of
surface position only and does not involve solid angles. The two point transport
intensity is defined by the following equation:
r
d = I(x x) = L(x, ) G(x, x) dA dA (2.49)
This is the flux flowing in the beam connecting dA to dA. Equation 2.48 can
be put in this form by multiplying both sides by G(x, x) dAdA which leads
to the following equation:

I(x x) = G(x, x) S fr(x x x) V(x, x) I(x x) dA (2.50)

Equation 2.50 defines the amount of light flowing from x to x and reflected to
x. Thus, it is sometimes referred to as the multipoint transport equation (see
Figure 2.16). The quantity
r r
fr(x x x) = fr(x, i ) (2.51)

is just a reparameterization of the BRDF.

There is one final step required to arrive at the full rendering equation, and
that is to account for all modes of light transport at a surface. In an environment
consisting only of opaque surfaces, the only other source of light is due to
emission from the surface.
r r
L(x, ) = Le(x, ) + S f (x) L(x, r ) G(x, x) V(x, x) dA
r (2.52)

where Le is the two point intensity of emitted light.

Radiosity and Realistic Image Synthesis 39

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 2. RENDERING CONCEPTS
2.6. THE RENDERING EQUATION

Figure 2.16: Three point transport geometry.

2.6.3 The Radiosity Equation

Finally, the rendering equation can be simplified given the radiosity assumption.
In radiosity, it is assumed that all surfaces in the environment are Lambertian
diffuse reflector. Thus, the BRDF is independent of the incoming and outgoing
directions and can be taken out of the integral.

L(x x) = Le(x x) + fr(x) S L(x x) G(x, x) V(x, x) dA

= Le(x x) + (x)

S L(x x) G(x, x) V(x, x) dA
(2.53)

More importantly, the outgoing radiance from a Lambertian surface is the same
in all directions and in fact equals the radiosity B divided by . This leads to
even more dramatic simplifications.4
G(x, x ) V(x, x )
B(x) = E(x) + (x)
S B( x )

dA (2.54)

The rendering equation expresses the conservation of light energy at all

points in space. The key feature of such an integral equation is that the quantity
to be computedin this case, the radiance or radiosityappears on the left-hand
side as well as under an integral on the right-hand side. For this reason, integral
equations are notoriously difficult to solve. They very rarely have closed-form
analytic solutions, and numerical methods must be used.

4
Note the switch in notation: E is the energy per unit area emitted by the surface,
or . In addition, for clarity in the following chapters, the geometric term G(x, x) will
absorb the visibility term and the p in the denominator.

Radiosity and Realistic Image Synthesis 40

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION

Chapter 3

Discretizing the Radiosity

Equation

3.1 The Radiosity Equation

The radiosity equation was derived at the end of Chapter 2 from the render-
ing equation under the assumption that all surfaces (and light sources) exhibit
Lambertian diffuse reflection (emission). Repeating the radiosity equation 2.54:

B(x) = E(x) + r(x) s B(x) G(x, x)dA (3.1)

where the geometric term, G(x, x), now includes the visibility term, V(x, x),
and division by . (A complete table of the mathematical terms used in this
chapter is provided in Tables 3.l and 3.2.)
The radiosity, B(x), describes an arbitrary scalar function across the sur-
faces (i.e., the radiosity function defines a single value at each location on a
surface).1 The potential complexity of the radiosity function is suggested by
Figure 3.1, where the radiosity function across a partially shadowed polygon is
plotted as a surface. The radiosity function is piecewise smooth, that is, it is
continuous in all derivatives within regions bounded by discontinuities in value
or derivatives. These characteristics will be discussed in much greater detail in
chapters 6 and 8.
The dimension of the function space of the radiosity function, B(x), is
infinite (for a discussion of function spaces, refer to the box on page 45). This
means that solving the radiosity equation for a point x on a surface does not

1
A full solution to the radiosity problem must also take into account the distribution of
energy across the visible spectrum (i.e., the color of the light). Assuming that the wave-
length of light is not changed by interaction with surfaces (i.e., ignoring fluorescence),
independent radiosity equations differing only in the reflectivities, r, can be formed and
solved for each of a small number of wavelengths or color bands. The selection of these
sample wavelengths and the reconstruction of colors suitable for display are discussed in
Chapter 9. Elsewhere in the book, the radiosity problem will be discussed in terms of an
achromatic (i.e., black, gray, white) world.

Radiosity and Realistic Image Synthesis 41

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.2 MAKING IMAGE SYNTHESIS TRACTABLE

S surfaces, the domain of the radiosity (and other) functions

n the number of basis functions or nodes
i, j indices into vectors and arrays
x, x two points in the domain S, e.g., (x, y, z), (x, y, z)
xi a specific point in S, the location of the ith node
A, Ai area, area of element i
dA a differential area at x, i.e., dx dy
B(x) radiosity function
B column vector of values. (B1, B2, ..., Bi, ...., Bn )T
(x) approximate radiosity function
E(x) emission function
E column vector of values, (E1, E2, ..., Ei, ...., En,)T
(x) diffuse reflectivity
i diffuse reflectivity at xi or of element i
(x) error function
r(x) residual function
K matrix of interaction coefficients

Table 3.1: Table of Terms

determine the radiosity at an immediately neighboring location. As a result, a

full and exact solution to the radiosity equation requires either finding the exact
functional form of the radiosity across each surface or computing radiosity values
for an infinite number of surface points.
The first step in constructing a practical global illumination algorithm is thus
to reformulate the problem as one that can be solved for a finite set of unknown
values, which can then be used to construct an approximate solution. This is
the topic of the present chapter, in which the finite element method will be used
to derive a linear system of equations that can be solved for discrete unknowns.
These unknowns then define an approximation to the radiosity function.

3.2 Making Image Synthesis Tractable

A great deal of research has been conducted on solving integral equations such
as the radiosity equation. A good source for this work is [69]. The finite element
approach used in radiosity methods is one of two distinct approaches that have
been taken in image synthesis to solving the global illumination problem.
One basic approach is the use of Monte Carlo or quasi-Monte Carlo ray

Radiosity and Realistic Image Synthesis 42

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.2 MAKING IMAGE SYNTHESIS TRACTABLE

Kij ijth term of K

F matrix of form factors
M matrix of basis interactions, often the identity, I
P diagonal matrix of element reflectivities
Ni(x) the ith basis function
Ni(xj) the ith basis function evaluated at xj
G(x, x) function of geometric relationship between x and x
Gij geometric relationship between points in elements i and j
Gi geometric function of a point and a direction
V(x, x) visibility relationship between x and xi
Vij visibility relationship between points in elements i and j
q angle between normal at x and vector from x to x
q angle between normal at x and vector from x to x
the hemispherical solid angle around the surface normal
r |x x|, the distance from x to x
ij Kronecker delta, 1 if i = j, 0 otherwise
u,v parametric coordinates

Table 3.2: Table of Terms (cont.)

tracing to solve the rendering equation for locations and directions determined
by the view specification. This approach solves the rendering equation for only
those surface locations that are visible in the image and for only those directions
that lead back to the eye. If interreflection is limited to the ideal (mirror)
specular component and light sources are limited to point lights, Monte Carlo
evaluation of the rendering equation is essentially equivalent to classical ray
tracing. Ray tracing and Monte Carlo methods for image synthesis have been
extensively investigated [64, 135, 198, 215]. Basic Monte Carlo integration will
be discussed in Chapter 4.
Because the solution is limited by the view, ray tracing is often said to
provide a view-dependent solution, although this is somewhat misleading in that
it implies that the radiance itself is dependent on the view, which is not the case.
The term view-dependent refers only to the use of the view to limit the set of
locations and directions for which the radiance is computed. When a new view
is specified, previously computed radiances are still valid but are often no longer
relevant, and the rendering equation must be solved for a new set of locations
and directions.
The radiosity method is based on the other basic approach to solving the

Radiosity and Realistic Image Synthesis 43

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.2 MAKING IMAGE SYNTHESIS TRACTABLE

Figure 3.1: An Image; the shadow on the floor; and the radiosity function across
the floor plotted as a surface.

global illumination problem, which is to compute and store an approximation

of the radiance function across the surfaces of the scene in object space, often
in the form of radiance values at discrete locations in the environment. During
a final rendering step, the shading of visible surfaces is derived from this ap-
proximation as needed. Because the view is not used to limit the scope of the
solution, a change in the view does not require recomputing the solution, and
this approach is often referred to as view-independent. Figure 3.2 shows both
approaches. Combinations of view-dependent and view-independent approaches
are also possible, and examples will be discussed in Chapters 9 and 10.

Radiosity and Realistic Image Synthesis 44

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.2 MAKING IMAGE SYNTHESIS TRACTABLE

Function Spaces

A function space defines a set of possible functions. For example, the ra-
diosity function is in the space of L2 functions over some finite domain S
(e.g., the surfaces). The L2 functions space is defined to contain all functions
f(x) such that
s | f (x) |2 dx <

In other words, the function is finite everywhere in the domain (except pos-
sibly in infinitesimal regions). Another function space would be the space
of all functions that are piecewise constant over integer intervals between 0
and 100. This function space can be defined as a linear sum of a box basis
functions:
99
f (x) = f
i=0
i Ni (x)

where the basis Ni(x) = 1 for i < x i + 1, and 0 elsewhere. The

coefficients, fi, describe the height of the function (step) in the ith interval.
The dimension of a function space is the number of discrete values required
to describe the function fully. Thus the dimension of the L2 function space
is infinite since the function can take on any form. In contrast, the piecewise
constant function space above is a finite function space since exactly 100
numbers (the coefficients fi) fully define the function.
Since any function in the piecewise constant function space is also in L2,
this finite function space is a subspace of L2.
The projection (or more precisely the orthogonal projection) of a general
function f(x) into the subspace (or basis) defined by basis functions Ni(x)
involves selecting the coefficients or weights, fi, so as to minimize the l2
norm (see section 6.1) of the difference between the original function, f(x),
and f (x) , (i.e., min || f (x) i fi Ni (x) ||2).

Computing and storing an object space approximation is particularly straight-

forward in the case of the radiosity equation, since for Lambertian diffuse reflec-
tion the radiance is independent of direction and is simply a constant multiple
of the scalar radiosity value at any given point on a surface. This means that the
approximation needs to define only a single value for each surface location. The
radiosity method was the first application of this approach to image synthesis.

Radiosity and Realistic Image Synthesis 45

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.3 THE RADIOSITY APPROACH

Figure 3.2: View-dependent and view-independent approaches to computing

global illumination.

3.3 The Radiosity Approach

As shown by Heckbert and Winget [125], the radiosity method as developed
in the field of radiative heat transfer can be viewed as a simple finite element
formulation.2 Finite element methods represent a complicated function by sub-
dividing the domain into elements over which the function is approximated by
a linear sum of simple basis functions also sometimes called shape functions
[38, 273]. The contribution of each basis function to the approximation is de-
termined by values computed at a finite number of locations, or nodes. These
nodal values become the unknowns in a linear system of equations, which can
be solved using a variety of techniques.
The basic steps to formulating and solving a generic radiosity problem are
outlined here (corresponding to the flow diagram in Figure 3.3):

1. Subdivide the surfaces into elements.

2. Select locations or nodes on the elements at which to determine discrete

radiosity values. These will ultimately become the finite set of unknowns
in the linear system of equations.

2
In the case of the radiosity function, whose domain is the surfaces of the environment
but where the interaction occurs across the space between surfaces, boundary element
methods are specifically relevant. The basic approximation method in finite and boundary
element methods is the same.

Radiosity and Realistic Image Synthesis 46

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.3 THE RADIOSITY APPROACH

Figure 3.3: Flow chart of traditional radiosity program.

3. Assume a simple functional form for the variation of radiosity across an

element. This is accomplished by associating a basis function with each
node. The basis functions are typically simple polynomials with local
support (i.e., zero everywhere except in the neighborhood of the associated
node). When weighted by the nodal radiosity values, the basis functions
for a particular element sum to define an approximate radiosity at each
point on the element.

4. Select a finite error metric to minimize by projecting a residual function

onto a set of basis functions. This approximates the error as a finite
sum, thus casting the infinite dimension problem into a finite set of linear
equations.

5. Compute the coefficients of the linear system. These coefficients are pri-
marily based on the geometric relationships that determine the transport
of light between elements. The geometric relationships are called form
factors.

Radiosity and Realistic Image Synthesis 47

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.4 APPROXIMATING RADIOSITY ACROSS A SURFACE

6. Solve the resulting system of equations for the unknown nodal radiosity
values.

7. Reconstruct the approximate radiosity solution as the linear sum of the

basis functions weighted by the resulting nodal values from step 6.

8. Render an image, deriving shading values from the radiosity approxima-

tion as needed for the particular view. The approximation used during the
solution may be modified during this stage to meet the particular require-
ments of rendering.

Note that this is a generic radiosity method; many of the enhancements

to the radiosity method that will be discussed in later chapters of this book
complicate this simple flow by looping back to repeat earlier steps. It should
also be noted that steps 3 and 4, in particular, are stages in formulating the basic
solution approach and should not be interpreted as steps that might be explicitly
performed by an algorithm.
The goal of the first four steps in the previous outline is to derive a linear
system of equations that can be solved for a finite number of discrete unknowns.
The remainder of this chapter will expand the explanation of these steps to
provide techniques for obtaining this system of equations.

3.4 Approximating Radiosity across a Surface

The heart of the finite element method is the approximation of a complicated
function by a linear sum of a finite number n, of simpler basis functions, Ni (x),
each of which is nonzero over a limited region of the function domain. The
domain of the function (the surfaces, in the case of radiosity) is subdivided into
a mesh of elements, each of which has one or more nodes located at points, xi ,
inside or on the element boundary. The mesh of elements and nodes organizes
the basis functions.
Each node is associated with a basis function. Each basis function, although
defined over the full domain of the original unknown function, is chosen to have
a zero value everywhere except over a small region of the domain. This support
of the basis function is confined to the elements adjacent to the node (a node on
an element boundary may be adjacent to two or more elements). This limited
support implies that the coefficient of single basis will only effect a small range
of the overall function. The approximation within an element is thus determined
by summing only the few basis functions whose support overlaps the element,
with the contribution of each basis function weighted by a coefficient value, Bi ,

Radiosity and Realistic Image Synthesis 48

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.4 APPROXIMATING RADIOSITY ACROSS A SURFACE

at the associated node. The radiosity function, B(x), is thus approximated by

B (x), where
n
B(x) B(x)
= B N (x)
i=1
i i (3.2)

Evaluating B (x) at a particular point xj involves summing only those basis

functions with their support over the point.
There are many possible basis functions. Low-order polynomials are the
most common, including constant, linear, quadratic, and cubic functions. The
constant (or box) basis, which is often used in radiosity, is defined by

Ni (x) =
0 if x is outside element
(3.3)
1 if x is inside element
The number and placement of nodes within elements will depend on the order
of the polynomial. In the case of constant elements, defined by constant basis
functions, a single node located in the center of the element is commonly used.
The radiosity function across linear elements is defined by nodes and as-
sociated basis functions at the boundaries or corners (in two dimensions) of
the element. A simple example of function approximation using linear bases is
shown in Figure 3.4 for a function of a single variable. In this example, a linear
basis function is associated with each node, and the nodes are located at element
boundaries. The linear (or hat) basis function in one dimension is defined by

xx i 1
for xi 1 < x < xi
x x
x x
i i 1

Ni ( x) = x x i +1
for xi < x < xi +1
0
i +1 i (3.4)
otherwise


In Figure 3.4 the nodal values, Bi , are determined by simply evaluating the
function to be approximated at the node locations. In general, the function to be
approximated is not known, thus the coefficient values are computed to minimize
an error estimate or provide an overall best fit according to some criterion.
The generalization of finite element approximations to functions defined
over a two-dimensional domain, such as the radiosity function, is straightfor-
ward (see Figure 3.5). The elements are typically simple shapes, usually tri-
angles and/or convex quadrilaterals. Nodes for constant, linear, and quadratic
elements are shown in Figure 3.6. The approximation of a function for a single,
two-dimensional element using constant and linear basis functions is shown in
Figure 3.7. Note that the tensor product of the two one dimensional linear basis
functions results in a bilinear basis in two dimensions and is thus curved along
the diagonals.

Radiosity and Realistic Image Synthesis 49

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.4 APPROXIMATING RADIOSITY ACROSS A SURFACE

Figure 3.4: Finite element approximation of a function using linear basis func-
tions.

Radiosity and Realistic Image Synthesis 50

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.4 APPROXIMATING RADIOSITY ACROSS A SURFACE

Figure 3.5: Finite element approximation of the radiosity function.

Figure 3.6: Dividing the surface into elements with nodes.

Radiosity and Realistic Image Synthesis 51

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.4 APPROXIMATING RADIOSITY ACROSS A SURFACE

In two dimensions, the linear basis function, Ni (x), has the properties:

1 at node i
0...1
within adjacent elements
Ni (x) = (3.5)
0 at all other node points
0 outside adjacent elements
This definition ensures that the basis function associated with a particular node
is zero outside of the elements adjacent to the node, as well as at any other
nodes within those elements.

The variation of the linear basis function from a value of 1 at the node
to 0 across the element depends on the parametric mapping of a generic basis
defined on a standard element to the triangular or quadrilateral elements. In the
case of triangles, barycentric coordinates are often used to define the 2D linear
basis. In the case of quadrilaterals a tensor product of two one-dimensional
linear basis functions defines one quarter of a bilinear basis within each element.
For example, the value of the hat function is determined by the interpolation,
Ni (x) = (1 u)(1 v), where (u, v) is the (0, 0) to (1, 1) parametric coordinate
of the point x within the element with node i at the parametric origin (see
Figure 3.9). Parametric mappings are discussed briefly in section 3.8 and in
more detail in most finite element texts [273].

Elements constructed using other types of basis functions are also possible.
For example, cubic Hermite elements provide continuity of derivatives across
element boundaries by using basis functions that interpolate function derivatives
at the nodes as well as values. Given the radiosity values and their paramet-
ric derivatives at the four corners of a quadrilateral, one can derive a bicubic
variation of radiosity across the element. Hybrid sets of basis functions with
constant, linear, and higher-order functions can also be employed. Other poly-
nomial bases (e.g., Jacobi polynomial sets [270]) and non-polynomial functions
are other possible choices. For a given approximation accuracy, there is typi-
cally a tradeoff between the number of elements required and the complexity of
the basis functions.

To date, radiosity implementations have used constant basis functions almost

exclusively. Linear functions have been explored in one dimension by Heckbert
[122]. Max and Allison [163] have implemented radiosity using linear elements.
Lischinski et al. [154] have used quadratic elements in a radiosity solution. For
radiosity, as will be discussed at length in Chapter 9, elements of linear or higher
order have more commonly been used during the rendering stage [203].

Radiosity and Realistic Image Synthesis 52

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.5 ERROR METRICS

Figure 3.7: Basis functions.

3.5 Error Metrics

A set of discrete elements, with a total of n nodal values at n nodal points, xi ,
and n basis functions, Ni , defines a finite dimension function space (see box
on page 45). This finite function space has n degrees of freedom, the n nodal
(x) is everywhere defined by the finite set of the n coefficients,
values, Bi ,thus B
Bi.
Ideally, one would like the approximate radiosity solution B (x) to agree

Radiosity and Realistic Image Synthesis 53

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.5 ERROR METRICS

with the actual solution B(x) everywhere. This is not possible in general since

B(x) is restricted to the finite subspace of functions representable as the linear
sum of the selected basis functions. Given the choice of basis functions, the
goal is then to find the nodal values that generate a solution minimizing the error
according to some measure or metric.
The actual error, (x), is the difference between the approximate and exact
solutions, that is,
(x) = B(x) B(x)
(3.6)
Unfortunately, it is impossible to determine (x) directly since the exact solution
is not known.
An alternative characterization of the error is provided by the residual. If

B (x) is substituted for both occurrences of B in the original radiosity equation

B(x)
= E(x) + (x) s B(x) G(x, x) dA (3.7)

then the residual function is given by the difference of the left and right hand
ides

r(x) = B(x)
E(x) (x) s B(x) G(x, x) dA (3.8)

An exact solution will make r(x) zero everywhere, but this is unobtainable since,
as we have seen, B is restricted to the subspace of functions realizable using the
basis functions. Although B is restricted to a finite function space, the residual,
r, is not. However, a similar approach to that taken in approximating B can
be taken to project the residual into a finite function space. In this way, rather
than seek a solution technique that makes r(x) zero everywhere, the goal will
instead be to minimize r(x) according to a finite dimensional error metric.
The general approach is to choose n independent weighting functions, Wi(x),
each having local support, much like the earlier basis functions. The norm or
size of the residual is then approximated as a finite sum:3
n
r(x) =
i=1
< r(x), Wi (x) > (3.9)

This residual norm can be driven to zero by finding radiosity values such that
each of the n terms < r(x), Wi(x) > is equal to zero. As will be derived below,
3
The notation < f(x), g(x) > indicates the inner product of the two functions. The
inner product of two functions is analogous to the dot product of two vectors and is
defined to be
< f(x), g(x) > = s f(x) g(x) dx

Radiosity and Realistic Image Synthesis 54

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.5 ERROR METRICS

If < f(x), g(x) > = 0, then f(x) and g(x) are orthogonal.
setting each of these terms to zero results in a linear equation. The set of n
such linear equations (one for each weighting function) of the n nodal values,
Bi , can be solved simultaneously as discussed in Chapter 5.
This approach defines a general class of methods called weighted residual
methods. Two different choices for the weighting functions Wi(x) lead to two
contrasting formulations, point collocation and the Galerkin method, which will
be discussed in the following sections.

3.5.1 Point Collocation

The simplest set of weighting functions are delta functions:

Wi(x) = (x xi ) (3.10)

which are zero unless x is coincident with the node at xi .

Using these weighting functions, the norm of the residual defined by equa-
tions 3.8 and 3.9 is minimized (i.e., zero) when r(x) is exactly zero at all the
node points, xi (i.e., r(xi ) = 0, i). This clearly differs from requiring r(x) to
be zero everywhere, since r(x) is free to vary away from zero in between the
nodes. However, as the number of nodes increases this difference diminishes.
In general, the points at which the residual function is minimized are selected
to be located at the nodes used in approximating the radiosity function. This
technique is known as point collocation.
If there are n nodes, there is now a finite number of conditions that must be
met. These are captured by the n simultaneous linear equations, one for each
node i located at location xi :

B (xi) E(xi) (xi) s B (x) G(xi, x) dA = 0, i (3.11)

Note that the only change from equation 3.8 to equation 3.11 is that equation
3.11 is defined only at the n node locations, xi , as opposed to the entire surface
domain, S. Expanding B using equation 3.2 gives

s B N (x) G(x , x) dA = 0
n

Bj Nj (xi ) E(xi ) (xi) j j i (3.12)

j =1 j =1

Grouping terms and factoring Bj, which is independent of x, out of the integral
leaves
n

B j N j (x i ) (x i ) S N j (x ) G(x i , x ) dA' E(xi ) = 0 (3.13)
j =1
Equation 3.13 can be restated simply as the set of n linear equations

Radiosity and Realistic Image Synthesis 55

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.5 ERROR METRICS

j =1
Bj Kij E(xi ) = 0 (3.14)

or in matrix/vector form as

KB=E (3.15)

The coefficients Kij are given by

Kij = Nj(xi ) (xi ) s N (x) G(x , x) dA

j i (3.16)

These coefficients are independent of the nodal radiosities and are defined solely
by the geometry and material properties of the environment. The evaluation of
these coefficients will be the topic of the following chapter. However, it should
be noted here that for any particular term, Kij, the integral only needs to be
evaluated over the nonzero support of the jth basis function. Also, the term
Nj(xi ) is simply the Kronecker delta ij , (i.e., one if i = j and zero otherwise),
for constant and linear basis functions with local support limited to adjacent
elements. Equation 3.14 completes the objective of deriving a linear system of
equations that can be solved for the unknown nodal radiosities, Bi .

3.5.2 Galerkin Form of Weighted Residuals

The point collocation formulation defines the error metric to be zero when the
approximate residual is zero at the node points only. An alternative approach to
deriving a linear system of radiosity equations defines the approximate residual
to be zero when a set of weighted integrals of the residual are zero.
The weighting functions in this case are selected to have local support much
like the basis functions used to approximate the radiosity function. In this case
the weighted residual method seeks a solution for which there is a weighted
average zero residual over each small region of the domain. If n weighting
functions, Wi(x), are defined, a solution is found if

< Wi(x), r(x) > = s W (x) r(x) dA = 0,

i (3.17)

or expanding r(x) above with equation 3.8

0= s W (x) B (x) dA
i

sW (x) E(x) dA s W (x) (x) s B (x) G(x, x) dA dA

i i (3.18)

Radiosity and Realistic Image Synthesis 56

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.6 CONSTANT ELEMENT RADIOSITIES

The Galerkin formulation selects the same basis functions used to approx-
imate the radiosity function as the weighting functions, (i.e., Wi(x) = Ni(x)),
thus, equation 3.18 becomes,

0= s N (x) B (x) dA
i

s N (x) E(x) dA s N (x) (x) s

i i B (x) G(x, x) dA dA (3.19)

Finally, expanding and grouping terms as before results in:

n

B j S Ni (x) N j (x) dA S N (x) (x) S N (x ) G(x, x ) dA dA
i j
j =1

s E(x) Ni(x) dA = 0 (3.20)

The unknowns, Bj , have been isolated in this expression. There are n such
expressions, one for each node i and, as with point collocation, the linear system
of equations can be expressed as the matrix equation

KB = E (3.21)

The entries of K are given by

Kij = s N (x) N (x) dA s N (x) (x) s N (x) G(x, x) dAdA

i j i j (3.22)

and the entries in E by

Ei = s E(x) N (x) dA
i (3.23)

The Galerkin method has been explored for one-dimensional flatland radiosity
by Heckbert [122, 123].

3.6 Constant Element Radiosities

The previous two sections have taken two different approaches to achieving
the same basic objective, a linear system of equations for the discrete nodal
values. However, the relationship of these equations to the physical problem
of global illumination may still be somewhat obscure. The complexity in the

Radiosity and Realistic Image Synthesis 57

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.6 CONSTANT ELEMENT RADIOSITIES

above formulations is reduced if constant basis functions are used, with the
reflectivity and emission of each element assumed to be constant. This set
of assumptions has been made in most radiosity implementations to date and
provides the clearest physical intuition.
For constant basis functions, Ni , the Galerkin formulation can be rewritten
by performing the integration for each term in equation 3.20. First, using the
fact that the box basis functions have values of only 11 and 0, and do not overlap
with one another,
Ai if i = j
s Ni(x) Nj(x) dA = = ij Ai (3.24)
0 otherwise
where ij is the Kronecker delta, and Ai is the area of element i. Note that
the integration only has to be performed across elements i and j since the
corresponding basis functions are zero elsewhere. Likewise,

s E(x) N (x) dA = E A
i i i (3.25)

where Ei is the area average emission value for element i.

Finally, since the basis functions have unit value within the element, the basis
functions themselves can be dropped and it is possible to integrate explicitly over
element areas. Thus, assuming the reflectivity (x) is a constant i over Ai ,

s Ni(x) r(x) s Nj(x) G(x, x) dA dA = i

Ai Aj
G(x, x) dAj dAi

(3.26)
Making these substitutions into equation 3.20 results in, for all i:

n

B j ij Ai i
Ai Aj G (x, x ) dA j dAi Ei Ai = 0 (3.27)
j =1

Dividing through by Ai and moving the emission term to the right side gives


n 1
B j ij

i
Ai Ai Aj G (x, x ) dA j dAi = Ei

(3.28)
j =1
or
n
[
B j ij i Fij
j =1
] = E i (3.29)

Radiosity and Realistic Image Synthesis 58

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.6 CONSTANT ELEMENT RADIOSITIES

where Fij , called the form factor, is given by

Fij =
1
Ai
Ai Aj G (x, x ) dA j dAi (3.30)

The form factor represents the fraction of energy that leaves element i and arrives
directly at element j. Evaluating the form factor is the topic of Chapter 4.
As in previous formulations, the system of equations can be represented by
the matrix equation K B = E. As evident in equation 3.29, the entries in K are
given by
Kij = ij i Fij (3.31)

Rearranging equation 3.29 gives the classical radiosity equation:

n
Bi = Ei + i B
j =1
j Fij (3.32)

A more physically intuitive form is obtained by first reinserting the area terms,
Ai: n
Bi Ai = Ei Ai + i Bj Fij Ai
j =1
(3.33)

and then using the reciprocity relationship between form factors,

Fji Aj = Fij Ai (3.34)
to obtain n
Bi Ai = Ei Ai + i B j =1
j Fij Aj (3.35)

A physical interpretation of equation 3.35 is that the total power, Bi Ai

leaving an element depends on any light that it emits directly plus light that is
reflected. This reflected light depends, in turn, on the light leaving every element
in the environment, since some fraction of the light leaving every other surface
may arrive at the surface in question and be reflected back into the environment.
This fraction depends on the geometric relationship between the elements and
on the element reflectivities.
A radiosity equation of the form of equation 3.32 exists for each element.
The system of equations can be expanded into matrix form:

1 1 F1,1 1 F1,n B1 E1
F 2 F2,n B E
2 2,1 2 2

= (3.36)

n 1 Fn 1,1 Bn 1 En 1

n Fn,1 1 n Fn,n Bn En

Radiosity and Realistic Image Synthesis 59

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.7 HIGHER-ORDER BASIS FUNCTIONS

Solving this linear system is the topic of Chapter 5. The radiosity formulation
for constant elements that has just been derived will be used for most of the
remainder of the book.

3.7 Higher-Order Basis Functions

The use of linear or higher-order basis (and residual weighting) functions does
not affect the basic formulation of the last section. A similar set of linear
equations results, but the physical interpretation is slightly altered. As described,
the form factor, Fij , from constant basis functions represents the fraction of
energy leaving element i that arrives directly at element j. Generalizing to
any basis functions, Fij represents the weighted effect of the energy leaving
the region under the support of one basis function on the energy of the region
under the support of another basis function (see Figure 3.8). The differential
interaction between two differential areas is similar to that with constant bases,
except in this case it is weighted by the local values of the two basis functions
at each end of the interaction. Although the physical intuition is somewhat less
direct, this is still essentially a geometric term.
The general K matrix can be expressed as the difference of two matrices,
M P F. For constant basis functions, the M matrix is simply the identity ma-
trix and the F matrix is the form factor matrix (P is a diagonal matrix of element
reflectivities ). Mij captures the inner product of the two basis functions, Ni
and Nj:

Mij = < Ni(x), Nj(x) > = s Ni(x) Nj(x) dA (3.37)

For orthonormal bases, this results in an identity since the inner products will
be zero unless they are the same function. Put differently, the terms in M
are zero unless the basis functions overlap as, for example, for linear functions
associated with nodes belonging to the same element. Thus, the M matrix is
sparse. The nonzero area terms, in the case of constant bases, are divided out
and disappear in the classical radiosity formulation.
In general, however, it is not possible to divide through by the area terms a
with the constant basis functions. Instead, one must be satisfied with the general
form
K=MPF (3.38)

with M from equation 3.37, and the unnormalized

Fij = s Ni(x) s Nj(x) G(x, x) dA dA (3.39)

Radiosity and Realistic Image Synthesis 60

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.8 PARAMETRIC MAPPING TO A MASTER ELEMENT

Figure 3.8: Form factors between basis functions.

3.8 Parametric Mapping to a Master Element

3.8.1 Master Elements

The complexity of handling many different types of surfaces and basis functions
can be ameliorated to a great extent by developing a bookkeeping scheme based
on the use of generic master elements. This topic is only briefly described here.
The specific derivations of and the use of master elements is described in great
detail in most finite element books such as [23, 273].
The radiosity function has been defined over the domain of the two-dimen-
sional surfaces embedded in three dimensional space. The basis functions are
defined in terms of the shape and size of the elements over which they have
non-zero support. Thus, each basis function is unique in terms of its location,
size and shape. One would like to be able to perform certain computations
independent of the specific element under consideration. For this reason, it is
useful to develop the notion of a small set of standard master elements and as-
sociated basis functions to which all elements can be mapped. Generic methods

Radiosity and Realistic Image Synthesis 61

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.8 PARAMETRIC MAPPING TO A MASTER ELEMENT

Figure 3.9: Parametric mapping.

for interpolation, integration of form factors, and other operations can then be
developed to act on the standard elements in their local setting and the results
transferred to the specific element in the global environment. This eliminates a
great deal of awkwardness that may be associated with working directly in the
more general global setting of planes and curved surfaces in three-dimensional
space.
Although many master element shapes are possible, most applications use
triangles and quadrilaterals. The original elements may or may not be planar;
however, it is important that there is an invertible mapping from the world space
coordinates to the two-dimensional (u, v) parametric coordinates of the master
element (see Figure 3.9). Typically, the master element coordinates are defined
over the unit 0 to 1 interval or the two unit -1 to 1 interval.
In the case of planar elements, the (u, v) coordinates are typically derived
by establishing a coordinate system on the plane of the element. Surface patches
that are already defined in terms of parametric coordinates may require a further
scaling to the master element coordinate system.
It is also important to be able to determine the transformation of differential
areas to the parameterized space. This requires careful attention in the case of
parametrically defined polygons and curved surface patches.

3.8.2 Isoparametric Mapping

The above discussion has not relied on the actual form of the mapping functions
other than to specify that they should be one to one and invertible. (In practice
this means that polygonal elements must be convex.) Although many mapping

Radiosity and Realistic Image Synthesis 62

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.9 SUMMARY

functions are possible, commonly the basis functions themselves are used as the
mapping functions. This is called an isoparametric mapping.
In this case, both the radiosity function and the geometry are mapped in
the same way. For linear basis functions, the one-to-one mapping requirement
will permit the use of triangles and convex quadrilaterals in parametric space.
Quadratic basis functions will permit three- and four-sided elements bounded by
quadratic curves to be mapped to square master elements. A full set of element
types and mapping functions can be found in most finite element texts such as
[273]. In the case of constant basis functions a superparametric mapping is
often employed using barycentric and bilinear functions to map triangles and
quadrilaterals to equilateral triangular and square master elements.
The example of linear bases and a square master element ranging from (0,0)
to (1,1) results in four mapping functions for the four corner nodes,

1
N0,0 (u, v) = 4 (1 u) (1 v)

1
N1,0 (u, v) = 4 (v) (1 v)
(3.40)
1
N1,1 (u, v) = 4 (v) (v)

1
N0,1 (u, v) = 4 (u) (v)

These mapping functions can then be used to find the world coordinates corre-
sponding to a parametric location inside a master element, thus defining the shape
of the element in world coordinates. If the world coordinates of the four corners
of a quadrilateral element are (in counter-clockwise order) x1, x2, x3, and x4, then

4
x(u, v) = N
i =1
i (u, v)x i (3.41)

The inverse of this mapping function is described in section 9.2.1 in the context
of interpolating radiosity values. In a similar fashion, curved element geometries
can be defined with higher order basis functions.

3.9 Summary
This chapter has provided a mathematical framework for approximately solving
the radiosity integral through the solution of a system of linear equations. Given
this framework, many practical questions remain to be explored. What is a
good element subdivision and how can it be obtained automatically? How are
the coefficients of the K matrix to be computed numerically and how is the

Radiosity and Realistic Image Synthesis 63

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 3. DISCRETIZING THE RADIOSITY EQUATION
3.9 SUMMARY

accuracy of that computation to be controlled? What matrix techniques are

appropriate to solving this system of equations? How is the resulting solution
applied to the rendering of an actual image? Efficiency and accuracy will be
primary concerns throughout the discussion of these and other questions in the
following chapters.

Radiosity and Realistic Image Synthesis 64

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR

Chapter 4

The Form Factor

I. The Form Factor Integral

As shown in Chapter 3, the solution of the radiosity integral equation using

point collocation or Galerkin methods leads to a system of linear equations of
the form
KB = E (4.1)

Evaluating the terms of K is the major computational bottleneck of the radiosity

method. This chapter will address the computation of the entries in K, compar-
ing algorithms according to their efficiency and accuracy. Chapter 7 addresses
another approach to increased efficiency, that is, limiting the number of entries
in the operator that are actually computed.
The matrix K is defined as the difference of matrices:

K = M PF (4.2)

where M and P are diagonal in structure and defined by the selected basis
functions and element reflectivities, respectively. The form factors (entries in F)
make up the most important component of K. They are central to the radiosity
method and to understanding the propagation of light within an environment.
Simply put, the form factor represents the geometric relationship between two
surface areas, although, depending on the selection of basis functions, the form
factor will take on more and less intuitive meanings. In the common case of
constant basis functions, the form factor represents the fraction of light leaving
one element that arrives directly at another. Thus, in the case where one element
is a light source, the form factor itself represents the direct illumination of the
other element per unit of emissive power from the light. The form factor is
purely a function of geometry. It does not depend on the reflective or emissive
characteristics of the surfaces. It does depend on the distance and orientation of
the two elements, as well as the visibility relationship between them.

Radiosity and Realistic Image Synthesis 65

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.1. The Coefficients of K

The computation of each form factor and thus each Kij involves evaluating
an integral of a geometric kernel over the areas of support of two basis functions.1
Integrals of this type can sometimes be solved in a closed form, and this will
be the case for special geometric arrangements of the elements. These special-
case analytic formulae will be discussed; however, in general, there will be no
closed form solution for this integral. Thus the bulk of the chapter will explore
a variety of numerical algorithms designed specifically for this problem.

4.1 The Coefficients of K

As described in Chapter 3, the matrix K results from the projection of the

radiosity function onto a selected basis set followed by the projection of the
linear integral operator itself (or the residual) onto a basis.2 The matrix K
consists of the difference of two matrices, M PF, where M accounts for the
overlap (inner product) of neighboring basis functions, P is a diagonal matrix
of surface reflectivities, and F is a matrix of form factors.
The coefficients of the matrix F represent the weighted direct influence of
the radiosity of a region under the support of one basis function to the radiosity
of another such region. Repeating from the previous chapter, the coefficients of
F are given by

Fij = s N (x) s N (x) G(x, x) dA dA,

i j (4.3)

where Ni(x) and Nj(x) are the basis functions evaluated at x and x and G(x, x)
incorporates the geometric and visibility relationship between the two regions
as it affects radiant energy transport.
In the case of constant basis functions, Ni(x) and Nj(x) are both equal to
1 over the finite areas of the corresponding elements. In this case, the M matrix
is simply a diagonal matrix of element areas. Thus, each equation in the system
can be divided by the area of the corresponding element. Then M is simply the
identity and

Fij =
1

Ai Aj Gij dAj dAi (4.4)
Ai
where the geometric kernel is now denoted using the shorthand Gij . In this case,
the coefficients Fij take on a particularly intuitive physical meaning: Fij is the
fraction of light leaving element i that arrives directly at element j. The Fij in
1
One basis may be a point as in the case of point collocation.
2
The same basis set is used in Galerkin methods.

Radiosity and Realistic Image Synthesis 66

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.2 THE DIFFERENTIAL FORM FACTOR

Figure 4.1: Form factor geometry.

this case are commonly called form factors.3 We will use the term form factor
for both constant and higher order bases, although the physical interpretation is
somewhat less straightforward in the more general case.
The form factor is purely a function of the geometric relationship between
elements and thus does not depend on the reflectivity (color) of surfaces. This is
a potential advantage of the radiosity method, since it allows surface attributes
to be changed without repeating the expensive form factor computation.

4.2 The Differential Form Factor

The geometric kernel in equation 4.3, G(x, x), was derived in Chapter 2. It
corresponds to the form factor between differential areas dA and dA and is
given by
cos cos cos
dF dAdA = d = dA (4.5)
r(x, x ) 2
3
Form factor has emerged as the common term in image synthesis. In the field of
heat transfer, the terms configuration factor and shape factor also appear, among others.

Radiosity and Realistic Image Synthesis 67

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.2 THE DIFFERENTIAL FORM FACTOR

where r = |x x|2 is the distance between x and x, is the angle between

the normal to the surface at x and the direction from x to x, is angle between
the normal to the surface at x and the direction from x to x, and d is the
differential solid angle subtended by dA from dA. The geometry for these
terms is shown in Figure 4.1.
The fraction of light leaving dA that arrives at dA is proportional to the
solid angle d subtended by dA as viewed from dA. The form factor thus
depends inversely on the square of the distance between the areas and on the
cosine of the angle . The form factor also drops off as dA moves toward the
horizon of dA, according to the cosine of the angle between the normal at dA
and the line from dA to dA. This follows from the definition of Lambertian
diffuse reflectivity, which dictates that the energy per unit solid angle leaving
dA in a certain direction is constant per unit projected area. The in the
denominator originates with the transformation from radiance to radiosity see
section 2.4.4), and acts to normalize the form factor to integrate to unity over
the hemisphere.
The differential form factor is a smooth function, in the sense that the cosine
and 1/r2 factors are continuous in all derivatives as dA moves about over the
domain. However, a singularity occurs as r 0 when the two differential areas
meet at some point (x = x). This can cause practical difficulties in evaluating
the form factor between elements that intersect or share a common edge.
Equation 4.5 does not account for occlusion due to surfaces positioned be-
tween the two differential areas, which block the direct transfer of light. De-
termining occlusion is often the most computationally demanding aspect of the
form factor computation since determining the visibility between two points may
require testing all surfaces in the environment. Also, because changes in visi-
bility can cause perceptually important discontinuities in the radiosity function,
visibility must often be determined to a high accuracy.
Occlusion is accounted for formally by adding a visibility term, V(x, x),
to equation 4.5. V(x, x) takes the value 1 if dA is visible from dA, and 0
otherwise. With the incorporation of occlusion, the complete differential form
factor equation becomes
cos cos
F dAdA = G(x, x ) dA = V(x, x )dA (4.6)
r2
where G(x, x) captures all the geometric terms.
Finally, it is useful to note that the differential form factor from one differen-
tial area to another, FdAdA, is equal to the reverse form factor, FdAdA . This
is known as the reciprocity relationship. The reciprocity relationship for form
factors follows from the Helmholtz reciprocity principle discussed in Chapter 2.

Radiosity and Realistic Image Synthesis 68

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.3. THREE FORMULATIONS OF THE FORM FACTOR

Figure 4.2: Differential solid angle.

4.3 Three Formulations of the Form Factor

The full form factor integral for constant bases, equation 4.4, involves performing
an integration over the support of two basis functions (i.e., the two element
areas). Three different formulations of this form factor integral between elements
can be derived. Each is used in the algorithms that will be described in the
following sections.
The first formulation results from inserting the differential form factor (equa-
tion 4.6) into the form factor integral, equation 4.4. Thus, the element-to-element
form factor is the double area integral, with geometric kernel, Gij ,
cos i cos j
Fij =
1
Vij dAj dAi (4.7)
r2
Ai Aj
Ai
where Vij is the visibility term between differential areas dAi and dAj .
The second formulation results from replacing the inner integration over area
with an equivalent integration over the hemisphere around dAi . The differential
solid angle (see Figure 4.2) dwj from dAi to dAj is
cos j
d j = dAj (4.8)
r2
Thus, the inner integral of equation 4.7 can be rewritten over the hemisphere of
directions, , and over dAi , resulting in the area-hemisphere integral,

cos i
Fij =
1
Vij d j dAi (4.9)

Ai
Ai

Radiosity and Realistic Image Synthesis 69

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.4. COMPUTING THE FORM FACTOR

r
where Vij is now 1 if element j is visible from dAi in direction d j . The new
geometric kernel, Vij cos i/, will be denoted Gi. This alternate form of the
integral will be used in some algorithms to compute the form factors from dAi
to all elements at once.
If the elements are assumed to be fully visible to one another, a third vari-
ation of the form factor integral can be derived by applying Stokes theorem to
convert the double area integral into a double contour integral [143, 222]. The
result (not derived here) is

1
Fij = ln r dxi dx j + ln r dyi dy j + ln r dzi dz j (4.10)
2 Ai Ci C j
where Ci and Cj are the boundaries of elements i and j. This third form of the
integral will also be used by some algorithms when visibility can be determined

4.4 Computing the Form Factor

Both closed form analytic and numerical methods have been applied to solving
the form factor integral.4
4
An excerpt from Schrder and Hanrahan [206]: The history of computing the
amount of light impinging on a diffusely reflecting surface from some light source is
very long. A closed form expression for the form factor between a differential surface
element and a polygon had already been found by Lambert in 1760 [143]. Lambert
proceeded to derive the form factor for a number of special configurations among them
the form factor between two rectangles sharing a common edge with an angle of 90
degrees between them. He writes about the latter derivation:
Although this task appears very simple its solution is considerably more
knotted than one would expect. For it would be very easy to write down
the differential expression of fourth order which one would need to inte-
grate four fold; but the highly laborious computation would fill even the
most patient with disgust and drive them away from the task. The only
simplification which I was able to achieve was to reduce the expression
to a second order differential using [the formula for differential surface
element to polygon form factor (equation 4.15)] with which I was able to
perform the computation.
Lambert also formulates the reciprocity principle in his theorem 16 and uses form factor
algebra to compute unknown factors from known ones. The first use of Stokes the-
orem [224] to solve for the form factor between two arbitrary surfaces can be found
in a book by Herman in 1900 [126]. Through two applications of Stokes theorem he
reduces the form factor between two arbitrary surfaces to a double contour integral. He
uses this result to give the form factor for two parallel quadrilaterals in an exercise. A
similar derivation can be found in an article by Fock in 1924 [83]. Fock proceeds by
applying the formulation to elliptical disks for which he derives a closed form solution.
In 1936 Moon [169] aware of Focks works, derives closed form solutions for a number

Radiosity and Realistic Image Synthesis 70

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.4. COMPUTING THE FORM FACTOR

Figure 4.3: A taxonomy of form factor algorithms.

A taxonomy of form factor computation methods is shown in Figure 4.3.

The discussion of these methods will begin with an examination of closed form
(analytic) solutions. Although there is no closed form solution to the general
form factor integral, if the two elements are assumed (or can be determined) to
be fully visible to each other, there are some useful analytic formulae.

of specialized configurations. In the same year Gershun [93] puts various photometric
quantities on a vector calculus footing and gives an especially elegant derivation of the
double contour integration using differential forms. Sparrow [221] in 1963 used the dou-
ble contour form to derive closed form solutions for the case of parallel disks and parallel
quadrilaterals. However none of these sources or any since that we are aware of has
iven a closed form solution of the form factor between two general polygons.

Radiosity and Realistic Image Synthesis 71

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.5 FORMULAE FOR SIMPLE SHAPES

II. Closed Form Solutions for the Form Factor

4.5 Formulae for Simple Shapes

Analytic formulae for specialized geometries can be found in the appendices of

most radiative heat transfer texts [131, 155, 216, 222]. The formulae given in
Figure 4.4 for opposing and perpendicular rectangles are typical. Although the
geometries are simple and visibility is not an issue, the analytic formulae ale far
from straightforward.
Analytic formulae are often used in conjunction with form factor algebra,
which allows the form factors for the union or difference of simple areas to be
computed from the form factors to the individual areas. As shown in Figure 4.5,
the form factor from element i to element j plus the form factor from element i
to element k must be equal to the form factor from element i to a new element
made up of the union of j and k. Similarly, the form factor from the union of
j and k to element i is the area average of the two individual elements. Thus,
if an element shape can be decomposed into simple shapes for which analytic
form factors are known, the form factor algebra can be used to determine the
full form factor.
Closed form formulae are also available for a differential area to various
finite geometries. An example is shown in Figure 4.6 for a parallel, axially
aligned disk. The differential area to finite area form factor arises naturally in
point collocation, where the radiosity equation is evaluated at element nodes. It
is also used in some numeric algorithms, since the outer integral in equations
4.7 and 4.9 can be performed numerically by evaluating the inner integral at one
or more locations on region i and averaging the result.

4.6 Differential Area to Convex Polygon

The analytic formula for the form factor from a differential area to a polygon is
particularly useful, since polygonal models are often encountered in image syn-
thesis [19, 130, 143, 175]. The geometry for this formula is given in Figure 4.7.
The form factor is computed as a sum around the perimeter of the polygon:
n

cos
1
FdAi A j = (4. 15)
2
i i
i =1

or equivalently,
n

N
1
FdAi A j = (Ri R(i +1)%n ) (4.16)
2
i i
i =1

Radiosity and Realistic Image Synthesis 72

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.6 DIFFERENTIAL AREA TO CONVEX POLYGON

Figure 4.4: Analytic form factors between rectangles.

where n is the number of sides on the polygon, bi is the angle between Ri and
R(i+1)%n in radians, ai is the angle between the plane of differential area dAi
and the triangle formed by dAi and the ith edge of the polygon, and Ni is the
unit normal to dAi.

This formula does not take into account occlusion. However, in conjunction
with the appropriate visibility algorithm, the form factor algebra can be used
to compute an exact result. In an algorithm by Nishita and Nakamae [175],
the form factor to the original polygon is first computed, ignoring occlusion.

Radiosity and Realistic Image Synthesis 73

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.7 GENERAL POLYGON TO POLYGON

F j,i A j + F k,i Ak
Fi,( j + k ) = Fi, j + Fi,k F( j + k ),i = A j + Ak (4.13)

Figure 4.5: Form factor algebra.

r2
Fij = (4.14)
h2 + r 2

Figure 4.6: Analytic form factor from point to disk.

Other polygons in the scene are clipped against this polygon (as viewed from
the differential area) and against each other. The form factor to each of the
clipped polygons is computed and subtracted from the unoccluded form factor,
giving the form factor to the visible portion of the original polygon.

4.7 General Polygon to Polygon

Schrder and Hanrahan give a closed form solution for general polygon-to-
polygon form factors, ignoring occlusion [206]. The formulation is non-elemen-
tary, as it is based on the dilogarithm [151] arising in the integration of the
contour integral form of the form factor, equation 4.10. The specifics of the
closed form solution are not repeated here as they involve a long series of
complex terms.

Radiosity and Realistic Image Synthesis 74

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.7. GENERAL POLYGON TO POLYGON

Figure 4.7: Differential area to polygon form factor.

III. Numerical Solutions for the Form Factor

Closed form analytic formulae do not lend themselves to the direct evaluation of
form factors between complex shapes or where occlusion is a factor. For more
general situations, numerical approaches are required to approximate the form
factor.
Numerical methods of evaluating integrals (known as quadrature methods)
generally involve sampling the kernel at various points and approximating the
integral as a weighted sum of the kernel evaluated at the sample points. In
general, the more sample points selected, the more accurate the approximation.
However, the cost of approximating the integral is directly related to the num-
ber of kernel evaluations required (each of which generally requires a visibility
calculation). Thus, the goal in developing a numerical algorithm (or quadrature
rule) to solve the form factor integral is to get the most accuracy with the fewest
(and/or cheapest) kernel evaluations.
There are a number of choices to make in designing a quadrature rule for
form factor evaluation. First, in the case of constant basis functions, one can

Radiosity and Realistic Image Synthesis 75

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.8. NUMERICAL INTEGRATION IN GENERAL

choose to evaluate any of the different formulations of the form factor integral
given in section 4.3. The area-area form (equation 4.7) and the area-hemisphere
form (equation 4.9) are equivalent, and the contour form (equation 4.10) is also
a suitable choice if one can determine a priori that the elements are fully visible
to one another.
In addition, there is not just a single form factor to be evaluated, but rather
a matrix of form factors. Each entry of a row (or column) of the form factor
matrix shares a common element as the domain of integration, thus one may
take advantage of this coherence by simultaneously solving for a complete row
and/or column of form factors.
Finally, one is free to a great extent to choose where to sample the kernel
and how to weight the samples so long as it can be shown that as the number
of samples increases, the weighted sum converges to the true solution.
After a brief discussion of the general topic of numerical integration, the
chapter will proceed with a description of a variety of numerical algorithms that
have been developed for form factor evaluation. Other surveys of form factor
algorithms such as [78, 187] provide additional insights. The approaches are
broadly classified according to which form of the form factor integral they use,
how the sample points are selected and weighted, and in what order the form
factors are evaluated (e.g., one at a time or a row at a time).

4.8 Numerical Integration in General

Generically, quadrature rules approximate some integral H with kernel h(x)

over the domain X as a weighted sum :

H = X h(x) dx H = w h(x )
k =1
k k (4.17)

One is free to use any information available about h(x) in choosing the quadra-
ture points, xk .
Normally, one would like to make n as small as possible to limit the cost
incurred in the evaluation of the h(xk ), without compromising accuracy. The
simplest methods, like the trapezoidal rule or Simpsons rule, sample the domain
at evenly spaced intervals, evaluate the kernel at these points, and sum the results
weighted by the size of the interval. Clearly, as the number of samples increases,
and the size of the intervals decreases the approximation approaches the true
integral.

Radiosity and Realistic Image Synthesis 76

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.8. NUMERICAL INTEGRATION IN GENERAL

4.8.1 Gaussian Quadrature

Simple methods like the trapezoidal rule ignore any knowledge of the integrand.
A more efficient selection of quadrature points and weights can be made given
the available knowledge about the nature of the integrand h(x). For example, if
one knows h(x) is constant across the limits of the integral, then one quadrature
point anywhere is sufficient and the weight is simply the difference of the upper
and lower limits of integration (in our case, the area of the elements). In general,
the smoother the integrand, the fewer quadrature points required to approximate
the integral to within some given error.
In one dimension, Gaussian quadrature methods [185] can be used to evalu-
ate exactly integrals of polynomials up to order 2n + 1 with n carefully selected
points and proper weights. The theory behind this observation is quite elegant
[226]. The specific quadrature points and associated weights are tabulated and
can be found in most numerical methods books or can be computed through re-
currence relationships. Extensions to multiple dimensions as in the form factor
problem are possible but can be difficult due to the exponential growth in the
number of quadrature points required with respect to dimension.

4.8.2 Quadrature Points and the Form Factor Integral

In the double area integral form of the form factor (equation 4.7), the quadrature
point is now defined in the four-dimensional space, R2 3 R2, of the combined
elements.5 In other words, a quadrature point represents the selection of a pair
of 2D points, x and x, in elements i and j at which to evaluate the integrand.
Similarly, in the area-hemisphere form (equation 4.9), a quadrature point is in
r
the space R2 3 S2, that is, a point (x, ) is in the combined space of an element
area and the hemisphere of directions above that point.

4.8.3 Monte Carlo Methods

Monte Carlo methods are a family of quadrature methods of very general appli-
cability, often used for difficult equations where other methods are impractical
or unavailable. Monte Carlo techniques use random numbers to select sample
locations in the domain of integration at which to evaluate the integrand. 6 The
integral is then taken to be a weighted average of the kernel evaluation at sam-
ple points. The weights associated with each sample evaluation depend on how
5
R is the space of the real number line and S is the space of directions on a circle.
Thus R2 corresponds to a plane and S2 corresponds to the surface of a sphere.
6
Quasi-random distributions such as the Poisson disk may also be used. The sam-
ples in a quasi-random distribution are not completely independent but have statistical
properties that allow them to be used in place of random samples. See for example
discussions by Shirley in [212].

Radiosity and Realistic Image Synthesis 77

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.8. NUMERICAL INTEGRATION IN GENERAL

the samples are selected. If, for example, samples are evenly distributed in the
domain, then the weights are simply for n samples with A the size of the
domain of integration.
One would like, however, to make as few evaluations of the integrand as
possible to reach an answer with some level of confidence of being within some
error bound. This process can be enhanced by taking advantage of whatever
knowledge one has of the integrand. For example, if the integral H to be
evaluated has a kernel h(x) that can be factored, h(x) = f(x) g(x), where
g(x) is a simple known positive function (g(x) > 0 for x X), then the
integral can be written
H = X f(x) g(x) dx (4. 18)

In this case, one would like to find a distribution of samples xk in X that mimics
the known function, g(x). This method of selecting samples is called importance
sampling, since more samples are taken where g(x) is large (important) and
fewer are taken where g(x) is small. More formally,

H = X f(x) g(x) dx = X f(x) Gp(x) dx (4.19)

where
G = X g(x)dx and p(x) =
g( x )
G
(4.20)

p(x) is essentially a normalized g(x) (i.e., p( x) = 1) and is thus a prob-
ability density function. The cumulative density function P(x) can be defined
as

x
P(x) = p(x) dx

(4.21)

Loosely speaking, if p(x) represents the odds of picking x, then P(x) represents
the odds of picking some number less than or equal to x. If the inverse of the
function P(x) is P1(x), then the Monte Carlo evaluation for the approximation
H H is simply

H = 0;
for ( k = 1 to n) { for n samples
choose ; randomly in the interval from 0 to 1
x = P1 () ; x will be chosen with probability p(x)
H = H + f(x) ; sum the sample values of f(x)
}
H = H * Gn ; normalize by G and divide by n samples

Radiosity and Realistic Image Synthesis 78

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

This type of importance sampling will be used, for example, to distribute

samples according to the cosine in the form factor integral to reduce the number
of expensive visibility evaluations. Explicit inverses for P(x) may be difficult
to derive, in which case analog procedures, or precomputed lookup tables may
be useful. Malleys method, discussed in section 4.9.5, is an example of this
approach.
Monte Carlo integration in general is a broad topic, to which numerous
textbooks and articles are devoted. Monte Carlo integration for image synthesis
is discussed in greater detail in [71, 135, 147, 215]

4.9 Evaluating the Inner Integral

The area-area and the area-hemisphere forms of the double integral both share
the same outer integral over element i. Thus, one can separate the question
of evaluating the full integral into two parts: first, the evaluation of the inner
integral from some dAi , and second, the selection of one or more sample points
on element i at which to evaluate the inner integral. If points on element i are
chosen with an even distribution then the full integral is the average of the inner
integral evaluations.
In fact, many implementations use only one representative point on element
i, in essence making the following assumption (valid also for the area-area
form):

Fij =
1

Ai Gi d dAi Gi d at sample point xi (4.22)
Ai

This assumes that the inner integral is constant across element i, which may
be reasonable when the distance (squared) is much greater than the area of
element j. However, changes in visibility between elements i and j also affect
the validity of this approximation. Whether or not this assumption is used, or
the inner integral is evaluated at many sample points in element i, one is left
with the problem of evaluating the inner integral. This will be the topic of the
following sections.

4.9.1 Hemisphere Sampling Algorithms

The use of the hemispherical integral provides a direct basis for evaluating the
form factor from a differential area, dAi , to all elements simultaneously. The
geometric kernel, Gi , is given by

cos i
Gi = Vij (4.23)

Radiosity and Realistic Image Synthesis 79

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

r
The visibility term, Vij , which is 1 if element j is visible in direction d , is
the computationally expensive part of the kernel evaluation. Determining the
visibility of an element j in a particular direction involves intersecting a ray
emanating from dAi with element j and with all surfaces that may block the
ray.
Since form factors from element i to all elements are needed eventually,
there is a clear advantage in performing the visibility calculation once and simply
summing a differential form factor into Fik , where element k is the element
r
seen in direction d from dAi. Thus, a single integration over the hemisphere
results in a full row of differential area-to-element form factors. Given this
observation, the question is how to sample the hemisphere of directions above
dAi , and how to perform the visibility calculations in the chosen directions.
Hemisphere sampling approaches can be principally differentiated by the
hidden surface algorithm they use. Certain algorithms provide efficiency by
exploiting coherence in the geometry of visible surfaces, while others provide
flexibility for stochastic, adaptive or nonuniform sampling. A geometric analog
to the differential area-to-area form factor equation is given first to provide some
useful intuition for the algorithms that follow.

4.9.2 Nusselt Analog

The Nusselt analog provides a geometric analog to the differential area-to-area
form factor equation 4.7 (ignoring the visibility factor). An imaginary unit
hemisphere is centered on the differential area, as in Figure 4.8. The element
is projected radially onto the hemisphere and then orthogonally down from the
hemisphere onto its base. The fraction of the base area covered by this projection
is equal to the form factor.
Why does this work? The area of the projection of the element on the unit
hemisphere is equal to the solid angle subtended by the element, by definition
of the solid angle, and thus accounts for the factor cos j /r2. The projection
down onto the base accounts for the cos i term, and the in the denominator
is the area of a unit circle.
In heat transfer applications, form factors have sometimes been computed
for complex shapes by evaluating the Nusselt analog photographically, using
a fisheye lens that effectively performs the same double projection. The area
covered by the object in the resulting photograph is measured manually to obtain
the form factor.

4.9.3 The Hemicube

The Nusselt analog illustrates the fact that elements covering the same projected
area on the hemisphere will have the same form factor, since they occupy the

Radiosity and Realistic Image Synthesis 80

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.8: Nusselt analog. The form factor from the differential area dAi to
element Aj is proportional to the area of the double projection onto the base of
the hemisphere.

same solid angle. Likewise, if an element is projected radially onto any inter-
mediate surface, as in Figure 4.9, the form factor for the projection will be the
same as for the element itself. This observation forms the basis for the hemicube
form factor algorithm, in which elements are projected onto the planar faces of
a half cube instead of onto a hemisphere [62].
A hemicube is placed around a differential area (see Figure 4.10), with
the hemicube faces subdivided into small grid cells. Each grid cell defines a
direction and a solid angle. A delta form factor, F, is computed for each
cell based on its size and orientation (see Figure 4.11). For this purpose, it is
convenient to consider a unit hemicube (i.e., with a height of 1, and a top face
2 3 2 units), although the size of the hemicube is arbitrary, since it is the solid
angles subtended by the grid cells that are of interest. The delta form factors are
precomputed and stored in a lookup table. Only one eighth of the delta form
factors need be stored, due to symmetry (one eighth of the top face and half of
one side face).
Each face of the hemicube defines a perspective projection, with the eye
point located at the differential area and a 90 viewing frustum.7 The form
factor to an element is then approximated by projecting the element onto the
7
The sides of the hemicube actually define the top half of a 90 frustum since the
bottom half falls below the horizon.

Radiosity and Realistic Image Synthesis 81

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.9: Areas with same form factor. Areas A, B, and C, have the same
form factor as Aj from dAi .

faces of the hemicube and summing the delta form factors of the grid cells
covered by the projection. The visibility problem of determining the closet
surface for a regular grid of cells is, of course, a familiar one in computer
graphics, since it is essential to producing a raster image. The hemicube uses
the Z-buffer algorithm [84], which is simple and efficient, and has the additional
advantage of wide availability in hardware graphics accelerators.

The only difference between rendering onto the hemicube and normal image
rendering is that in addition to a Z depth, an ID for the visible element is stored
at each grid cell, instead of a color (the result is often called an item buffer after
[260]). The distances are initialized to infinite and the identifiers to NONE.
Each element in the environment is projected onto the face of the hemicube one
at a time. If the distance to the element through each grid cell is less than what
is already recorded, the new smaller distance is recorded as well as the identifier
for the element. When all elements have been processed, each grid cell will
contain an identifier of the closest element. The grid cells are traversed and the
delta form factor for each cell is added to the form factor for the element whose

Radiosity and Realistic Image Synthesis 82

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.10: The hemicube.

ID is stored with that cell. The form factor to element j is thus

Fi, j = F
qj
q (4.26)

where q represents delta grid cells covered by element j. Pseudocode is supplied

in Figures 4.12 and 4.13. Hall [114] also provides a detailed pseudocode for the
hemicube algorithm.
The hemicube algorithm defines a specific form of quadrature for evaluating
the inner form factor integral, F dA , j . The directions on the hemisphere are
i
predetermined by the orientation and resolution of the grid imposed on the
hemicube. The weights associated with each quadrature point are precisely the
delta form factors, F, described above. The F, in turn, have been evaluated

Radiosity and Realistic Image Synthesis 83

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

1
FdAi = Ai (4.24)
x + zi2 + 1
2
i

zi
FdA j = Aj (4.25)
y + z 2j + 1
2
j

Figure 4.11: Delta form factors.

by a one-point quadrature from the center of the hemicube to the hemicube pixel
centers.
The efficiency of the hemicube derives from the incremental nature of the Z-
buffer algorithm, which exploits coherency in the projected geometry. However,
the inflexibility of the Z-buffer is also the source of many of the hemicubes
disadvantages.
The hemicube divides the domain of integration, the hemisphere, into dis-
crete, regularly spaced solid angles. When projected elements cover only a few
of the grid cells, aliasing can occur. The result is often a plaid-like variation in
shading as the hemicube subdivision beats against the element discretization (see
Figure 4.14). Very small elements may fall between grid cells and be missed
entirely. This is a particular problem for small, high-energy elements like light
emitters. Increasing the hemicube resolution can reduce the problem, but this
Radiosity and Realistic Image Synthesis 84
Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

/* Preprocess: determine delta form factors, given a resolution of the

hemicube. Note: resolution may vary for sides and top of hemicube. Also
Note: symmetry can be used to minimize the storage and computation of
delta form factors. */

/* Top */
dx = dy = 2.0/res;
x = dx/2.0;
A = 4.0/(res2);
for ( i = 1 to res ) {
y = dy/2.0;
for ( j = 1 to res ) {
F[i][j] = 1.0/( * (x2 + y2 + 1.0)2);
y = y + dy;
}
x = x + dx;
}
/* Side, Note: keep track of ordering for scan conversion below */
dx = dz = 2.0/res;
x = dx/2.0;
A = 4.0/(res2);
for ( i = 1 to res/2) {
z = dz/2.0;/* Note: z goes from bottom to top */
for ( j = 1 to res ) {
F[i][j] = z/( * (x2 + z2 + 1.0)2);
z = z + dz;
}
x = x + dx;
}

Figure 4.12: Pseudocode for hemicube form factor calculation.

is an inefficient solution, since it increases the effort applied to all elements,

irrespective of their contribution.8
Because scan conversion requires a uniform grid, the hemisphere subdivision
imposed by the hemicube is also inefficient. From an importance sampling point
of view, the hemisphere should be subdivided so that every grid cell has the same
delta form factor, to distribute the computational effort evenly. Because it uses
8
Typical hemicube resolutions for images in [62, 246] range from 32 by 32 to 1024
by 1024 on the top hemicube face.

Radiosity and Realistic Image Synthesis 85

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

/* For each element, determine form factor to each other element */

for ( i = 1 to num_elements ) {
initialize Fij = 0 for all j ;
initialize all hemicube grid cells to NULL element ID ;
initialize all hemicube grid cells to large Z value ;
place eye at center (sample point on element i) ;
/* scan convert and Z-buffer element projections */
for ( top and each side of hemicube ) {
Align view direction with top or side;
for ( j = 1 to num_elements ) {
Project elementj onto hemicube ;
for (each grid cell covered)
if ( Z distance < recorded Z ) grid cell ID = j;
}
for ( j = 1 to num_elements )
Fij = Fij + F of grid cells with ID = j;
}
}

Figure 4.13: Pseudocode for hemicube form factor calculation (cont.).

uniform grid, the hemicube spends as much time determining form factors close
to the horizon as it does near the normal direction.
Max [162] has investigated variations of the hemicube in which the reso-
lution of the top and sides are allowed to differ, the cubical shape can become
rectangular, and the sizes of the grid cells are allowed to vary with position. By
assuming a statistical distribution of the elements being projected, he derives
optimal resolutions, shapes, and grid cell spacings to reduce quadrature errors.
For example, a top face resolution about 40% higher than for the sides, and a
side height of approximately 70% of the width provides are found to minimize
the error for a given number of grid cells.
The advantage of the hemicube is that it determines form factors to all
elements from a single point. This can also be a liability if only one form factor
is required. In addition, computing the full area-hemisphere form factor requires
repeating the hemicube9 at a sufficient number of sample points on element i
to ensure the desired accuracy for elements that are relatively close together.
9
Rotating the hemicube for each selected sample point is useful for eliminating aliasing
artifact.

Radiosity and Realistic Image Synthesis 86

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.14: Hemicube Aliasing.

Increasing accuracy in this way will be inefficient, since element i will be close
to only a fraction of the other elements in the environment, and the effort of the
extra hemicube evaluations will be wasted for elements that are distant from i.
In spite of these limitations, the hemicube can be a useful and efficient means
of computing form factors, as long as its limitations are taken into account.
Baum et al. [19] provide an extensive discussion of the inaccuracies introduced
by the hemicube, along with useful error estimates. Comparing results for the
hemicube to analytic results for a variety of geometries, Baum et al. find that
elements must be separated by at least five element diameters for the relative
error in the form factor to drop below 2.5 percent. This result will naturally
depend on the particular geometries and hemicube parameters, but it provides a
useful rule of thumb.
A cube is, of course, not the only solid constructed of planar faces, and
other shapes might be used similarly to the hemicube in constructing faces for
projection and scan conversion. For example, Beran-Koehn and Pavicic [24, 25]
describe an algorithm similar to the hemicube based on a cubic tetrahedron.
Spencer [223] describes the use of a regularly subdivided hemisphere.

Radiosity and Realistic Image Synthesis 87

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.15: Single plane method.

4.9.4 Single-Plane Method

The single-plane form factor algorithm developed by Sillion [218] partially

addresses the inflexibility of the hemicube by replacing the Z-buffer with an
adaptive hidden surface algorithm based on Warnock [255]. Sillions algorithm
projects elements onto a single plane above the differential area dAi . Warnocks
[84] algorithm is used to subdivide the image plane adaptively into windows for
which the element visibility can be determined trivially. Sillions algorithm is
able to subdivide coarsely for large elements and finely for small elements, and
thus avoids some of the sampling problems of the hemicube. The delta-form
factors associated with regions on the projection plane can be precomputed,
similarly to delta form factors for the hemicube.

Sillions algorithm can also compute form factors that take into account
specular surfaces. Rays are cast through the corners of the region if the visible
element is specular and traced recursively until a diffuse element is encountered
(see Figure 4.15).

A single plane algorithm similar to Sillions is described by Recker [191],

who uses a Z-buffer for the hidden surface removal. Both Sillion and Recker
note that the single plane will miss elements near the horizon. However, these
elements will typically contribute very little to the overall radiosity, due to the
cosine dependence of the form factor. Of course, there is no reason, in principle,
why a Warnock-style hidden surface algorithm could not be applied to the full
hemicube.

Single plane algorithms are based on projection, like the hemicube, and thus
compute form factors from a single point. The single plane thus has the same
problems as the hemicube for computing full area-hemisphere form factors.

Radiosity and Realistic Image Synthesis 88

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.16: Malleys method.

4.9.5 Monte Carlo Ray Tracing

Ray tracing (as opposed to scan conversion and the Z-buffer) provides an ex-
tremely flexible basis for evaluating the visibility term in the numerical integra-
tion of the form factor equation. Because rays are cast independently they can
be shot to and from any distribution of points on the elements or directions in the
hemisphere. Nonuniform and adaptive sampling can be used to distribute com-
putational effort evenly. Rays can also be distributed stochastically, which can
render the effects of inadequate sampling less noticeable by converting aliasing
to noise.
In addition, ray tracing handles a wide variety of surface types, including
curved surfaces, and a number of efficiency schemes to accelerate ray intersec-
tions exist. A disadvantage of ray tracing is that the expense per quadrature
point will generally be higher since coherency from ray to ray is more difficult
to exploit than in scan conversion algorithms.
Ray casting provides an excellent basis for Monte Carlo integration of the
form factor equation over the hemisphere. In equation 4.9, the kernel contains
the factor cos i . Thus importance sampling can be performed by selecting
directions over the hemisphere with a sample density proportional to the cosine.
In this way, more effort will be expended where the form factor is largest.
Since the density of rays is then proportional to the differential form factor,
each sample will carry equal weight.
Malley describes a straightforward method for generating a set of sample

Radiosity and Realistic Image Synthesis 89

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

directions with the desired distribution [157]. Malleys method is essentially a

Monte Carlo evaluation of the Nusselt analog (see Figure 4.8) run in reverse. He
begins by generating a set of random points uniformly distributed in the circle 10
under the hemisphere (see Figure 4.16).11 To determine a direction to shoot
a ray, one of these points is projected vertically to intersect the hemisphere.
The ray is then directed radially out from the center of the hemisphere through
this projected point. Rays are shot in this manner for every sample point in the
circle. The number of times each element in the scene is hit by a ray is recorded.
The form factor is then given by the number of rays that hit a given element
divided by the total number of rays shot. Referring back to Nusselts analog,
the total number of rays shot is an estimate of the area of the circle covered by
the double projection. The fraction of the total rays that hit a given element thus
approximates the area of the projection of the element on the hemisphere base,
relative to the total area of the base. This fraction is equal to the form factor.
Maxwell also describes the computation of form factors with ray tracing [164].

4.9.6 Area Sampling Algorithms

The hemisphere sampling algorithms described in the previous sections are most
efficient when form factors to all elements from a single point must be computed
at once. Certain solution techniques (e.g., the progressive radiosity algorithm
described in the next chapter) require form factors between only a single pair of
elements at a time, thus the full hemisphere methods are inefficient. In this case,
the area-area formulation (equation 4.7) is a more efficient basis for algorithm
development.
For convenience, the equation for the form factor from a differential area i
to a finite element j is repeated

cos i cos j
FdAi A j = Vij dAj . (4.27)
r2
Aj

The integration can be performed by evaluating the kernel at randomly dis-

tributed points for a Monte Carlo solution. Wangs discussion of Monte Carlo
sampling of spherical and triangular luminaires in [248] contains much practical
information that can be applied to form factor integration.
10
A random set of points in a circle can be derived by generating two random numbers
between 0 and 1 to locate a point in the square surrounding the circle. If the point is in
the circle, use it; if not discard it. Independently generating a random angle and radius
will not result in a uniform distribution of points.
11
These points are only used to determine a direction, not to select a point to start a
ray. For an area-to-area computation, the ray origin can also be stochastically distributed
over the area of the element.

Radiosity and Realistic Image Synthesis 90

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.17: Numerical integration of form factor from differential area to finite
area.

Alternatively, the integration can be performed by subdividing the area uni-

formly. Wallace et al. subdivide Aj into a set of m smaller areas Ajk and
select a sample point at the center of each subarea (see Figure 4.17). The form
factor then becomes

m cos ik cos jk
F dA A =
i j

k=1 (r )
k 2
V(dAi , A jk ) A jk (4.28)

The equation is evaluated by shooting a ray from dAi to each delta area to
determine the visibility, V(dAi , Ajk ). The contributions of those delta areas
that are visible is then summed.
Equation 4.28 assumes that the subareas are reasonable approximations to
differential areas, which is the case only if Ajk << r2. Otherwise, Ajk
should be treated as a finite area. For example, each term of the summation
could evaluate the exact polygon form factor formula for the particular subarea,
as discussed in Tampieri in [230].
A less expensive alternative is to approximate Ajk by a finite disk of the
same area, as suggested by Wallace et al. [247]. The analytic formula for a point-
to-disk form factor can then be substituted into the summation of Equation 4.28.
The form factor from a differential area to a directly opposing disk of area Aj

Radiosity and Realistic Image Synthesis 91

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.18: Configuration for approximation of the form factor from a differ-
ential area to arbitrarily oriented disk.

is
Aj
FdAi A j = (4.29)
r + Aj
2

The effect of element orientation can be approximated by including the

cosines of the angle between the normal at each surface and the direction between
the source and the receiver (see Figure 4.18):

Aj cos i cos j
FdAi A j = (4.30)
r 2 + Aj
The form factor from a differential area to an element j approximated by a
set of m disks of area Aj /m is thus

m cos ik cos jk
F dA A = A j
i j

k=1 (r k ) 2 +
Aj V(dAi , A j ) (4.31)
m

The reciprocity relationship can also be used to approximate the form factor
from a finite area to a differential area through the ratio of the areas:

dAi cos i cos j

FA j dAi = FA j dAi = dAi (1.32)
Aj r 2 + Aj

The disk approximation breaks down when the distance from the disk to
the receiver is small relative to the size of the delta area, and visible artifacts
may result, as shown in Figure 4.19(a). An additional difficulty with uniform
subdivision of the element is that since a single ray is cast to each of the source
subdivisions, the element is essentially treated as several point lights as far

Radiosity and Realistic Image Synthesis 92

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.9. EVALUATING THE INNER INTEGRAL

Figure 4.19: (a) Artifacts introduced by the disk approximation. The receiving
surface consists of 30 by 30 elements. (b) Adaptive subdivision of the source
element for a typical node on the receiving element.

as visibility is concerned. As a result, the shadow boundary may appear as

a number of overlapping, sharp-edged shadows rather than a smoothly shaded
penumbra.
Both of these problems can be addressed by adaptively subdividing area
Aj . This is performed in a straightforward manner by subdividing the area
recursively until the resulting delta areas fulfill some criterion. The criterion
may be geometric (e.g., the delta area must be much less than r2) or based
on the energy received from the delta area. The result of adaptive element

Radiosity and Realistic Image Synthesis 93

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.10 FULL AREA-TO-AREA QUADRATURE

Figure 4.20: Monte Carlo area-to-area form factor.

subdivision is shown in Figure 4.19(b). Tampieri [230] provides a detailed

practical discussion of this approach, including pseudocode.

4.10 Full Area-to-Area Quadrature

Any of the analytic or numeric differential area-to-area form factor solutions

discussed so far can be used to approximate the full area to area form factor.
The differential area-to-area form factor is evaluated at one or more points on
element Ai and the result averaged. For example, the ray tracing algorithm
just described could be performed for several points on Ai . However, since
many rays connecting the two surfaces originate at the same points on Ai , this
approach samples Ai inefficiently. There are several more effective approaches,
including Monte Carlo integration, numerical solution of the contour integral
form, and hierarchical subdivision.

4.10.1 Monte Carlo Integration

The double area integral can be approximated more accurately by distribut-

ing the endpoints of the rays over Ai as well as Aj. In a Monte Carlo approach
ray endpoints on both elements would be distributed randomly, or according
to some quasi-random distribution like the Poisson disk. Pseudocode is for a

Radiosity and Realistic Image Synthesis 94

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.11. CONTOUR INTEGRAL FORMULATION

simple area-to-area Monte Carlo form factor algorithm is given in Figure 4.21
(the geometry is shown in Figure 4.20).
Fij = 0
for k = 1 to n
randomly select point xi on element i or use stratified sample
randomly select point xj on element j or use stratified sample
determine visibility between xi and xj
if visible
compute r2 = (xi xj)2
r r
compute cosj = rij Ni
r r
compute cosj = rij N j

compute F = cos icos j

p r 2 + Anj
if( F > 0 ) Fij = Fij + F
end if
end for
Fij = Fij * Aj
r r
where rij is the normalized vector from xi to xj, and N i is the unit normal to
element i at point xi (and vice versa for switching i and j).
Figure 4.21: Pseudocode for Monte Carlo area-to-area form factor computation.
One can do better in terms of fewer rays by sampling the elements nonuni-
formly and adaptively. An elegant solution for this decision-making process is
presented in Chapter 7.

4.11 Contour Integral Formulation

In the earliest work introducing the radiosity method to computer graphics, Goral
et al. [100] used Stokes, theorem to convert the double area integral into the
double contour integral of equation 4.10.
The contour integrals can be evaluated numerically by walking,, around
the contours of the pair of elements,12 evaluating the kernel at discrete points
and summing the values of the kernel at those points [100]. In fact, Goral
et al. use a three-point quadratic Gaussian quadrature (nine-point in 2D) along
12
Nishita and Nakamae point out that the contour integration approach can be used to
compute a single form factor to objects constructed of multiple non-coplanar surfaces.
The form factor computed for the contour of the composite object as viewed from the
differential area is equal to the sum of the form factors to each of the component surfaces,
since it is the solid angle subtended by the surfaces of the object that determines their
contribution to the form factor. This is a valuable observation that has not yet been taken
advantage of in radiosity implementations for image synthesis.

Radiosity and Realistic Image Synthesis 95

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.12 A SIMPLE TEST ENVIRONMENT

Figure 4.22: Simple test environment.

the boundaries. Care must be taken when the boundaries are shared between
elements, as ln(r) is undefined as r 0.
Equation 4.10 does not account for occlusion. It only the inner contour
integral is to be evaluated (in computing, a differential area-to-area form factor),
occlusion can be accounted for using a polygon clipping approach such as Nishita
and Nakamae s [175].

4.12 A Simple Test Environment

To provide a concrete illustration of some of the issues discussed in this chapter,

three numerical form factor algorithms have been tested on the simple two-
polygon environment shown in Figure 4.22. Results are shown in Figure 4.23.
The two polygons are parallel unit squares one unit distance apart. The analytic
form factor between them is approximately 0.1998.
Tests were run using the hemicube method, Malleys method for randomly
selecting directions, and the area-area Monte Carlo method. In each case, two
tests were run, (Test 1) from the center point only of element i, and (Test 2) from
a set of randomly selected points in element i. A series of 1000 runs was made
of each. The mean, standard deviation (box height in graph) and minimum and
maximum values (vertical lines) are displayed in the graphs.13 The horizontal
axis is given in terms of the resolution of the hemicube, the number of random
directions that fell in element j in Malleys method and the number of sample
points in element j in the Monte Carlo method. In Test 2, the same number
13
The hemicube method from the center of element i (Test 1) has no deviation since
it is a deterministic algorithm.

Radiosity and Realistic Image Synthesis 96

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.12 A SIMPLE TEST ENVIRONMENT

Figure 4.23: Test results on parallel elements. In Test1, only the center point
was chosen on element i and n points on element j. In Test 2 n points are chosen
on both elements. res is the resolution of the hemicube in both tests.

Radiosity and Realistic Image Synthesis 97

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.13. NONCONSTANT BASIS FUNCTIONS

was chosen for sample points in element i and the horizontal axis represents the
product of the number of sample points and directions.
All of the methods converged reasonably well in this case as the sampling
density (resolution, in the case of the hemicube) increased. Note that the form
factor from the center of element i to element j is approximately 0.2395, and
thus the solution always converged to about 20% over the actual value for Test
1. Also, note the single point on the graph for the hemicube with resolution 2
performed at 2 random points chosen on element i. Because of the low resolu-
tion and the fixed orientation of the hemicube with respect to the environment
coordinate frame, the form factor happens always to be the same, no matter
where the hemicube is position on element i. This extreme case highlights the
problems due to the regular sampling enforced by the hemicube that are en-
countered as the resolution becomes small relative to the projected area of the
elements.

4.13 Nonconstant Basis Functions

The discussion so far has been limited to evaluating form factors where con-
stant basis functions are used. This has been the dominant approximation used
in radiosity implementations to date. In this section we briefly introduce the
computation of form factors for higher order elements. However, this remains
largely an area for future development.
Recall from the last chapter that the coefficients of the linear operator K are
given by
Kij = Mij iFij (4.33)

For orthonormal (e.g., constant) bases the M matrix is simply the identity
after division by area terms. In the more general case, it represents the inner
product of the ith and jth basis functions:

Mij = s Ni(x) Nj(x) dA (4.34)

The integral will be nonzero only where the support of the two basis functions
overlaps. This integral can generally be evaluated analytically since the basis
functions are chosen to be simple polynomials.
Slightly reorganizing equation 4.3, the Fij are given by

Fij =
Ai Aj Ni (x) N j (x9) G(x, x9) dA9 dA (4.35)

The interpretation of the coefficients Fij is slightly less intuitive in this general
case. The integral is still over the support (Ai, Aj) of the two basis functions

Radiosity and Realistic Image Synthesis 98

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.13. NONCONSTANT BASIS FUNCTIONS

as before. The only difference is the inclusion of the basis functions in the
integrand. The form factor now represents the exchange of energy between
regions under the support of the basis functions, weighted by the local values
of basis functions.

4.13.1 The Hemicube for General Form Factors

Many numeric form factor algorithms can be generalized in a straightforward

manner to higher order basis functions. Wallaces ray casting algorithm, for
example, could evaluate the basis functions at each delta area and weight the
delta form factor accordingly. This is essentially the approach used by Tampieri
and Lischinski in [231].
Max and Allison [163] describe a hemicube algorithm for linear elements
that makes clever use of graphics hardware. Performing a hemicube for higher
order elements requires that the delta form factor for each grid cell be weighted
by the value of each basis function that is nonzero at the point on element j
seen through that particular grid cell.
Max and Allison use the Gouraud shading hardware in a graphics accelerator
to interpolate the basis function for triangular elements. Each of the three vertices
is given the maximum value for one of the r, g, b color channels. The faces of the
hemicube are set up as perspective views and the elements passed as polygons
to the hardware. The Gouraud shading hardware will then interpolate the vertex
colors to the hemicube grid cells. The value of each color channel gives the
value of the respective basis functions at that grid cell.
Noting that the three basis functions for a triangular element sum to 1.0
everywhere on the element interior, Max and Allison actually store colors for
only two vertices. The value of the third basis function can be obtained at any
grid cell using the identity N3(x) = 1.0 N1(x) N2(x). This frees up a
color channel, which when combined with the alpha channel provides room for
a 16-bit element ID.

4.13.2 Monte Carlo for General Form Factors

The Monte Carlo algorithm described earlier in section 4.10.1 for evaluating the
area-to-area form factor can also be generalized to higher order elements. The
algorithm is basically unchanged, except that each sample point on Ai and Aj
now contributes to the form factor for each of the nonzero basis functions at that
point, with the contribution weighted by the value of the basis function. A single
ray between two sample points x and x9 can be used to contribute to more than
one Fij if the basis functions overlap. For example, with linear, quadrilateral
elements, a single ray will typically result in 16 (4 i bases 3 4 j bases) nonzero

Radiosity and Realistic Image Synthesis 99

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.13. NONCONSTANT BASIS FUNCTIONS

Contributions to F17, F27, F18, and F28 will be nonzero

for the ray shown. Note: there would be 16 nonzero
contributions in 3D.

Figure 4.24: Monte Carlo for form factors with general basis functions.

Figure 4.25: Simple test environment with a singularity.

contributions to the K matrix (see Figure 4.24). The same samples can also be
used to evaluate the Mij numerically if desired.

4.13.3 Singularities in the Integrand

Although most of the algorithms described above will provide reasonable evalu-
ations of the form factor for most cases, they can suffer serious inadequacies in
the presence of singularities in the kernel. The singularity in the integrand for
equations 4.7 and 4.10 occurs when the two differential areas meet (i.e., r 0).
Potential difficulties are illustrated by an experimental comparison in which
three variations of numerical form factor algorithms were tested as before, but

Radiosity and Realistic Image Synthesis 100

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.13 NONCONSTANT BASIS FUNCTIONS

this time on two unit squares at right angles with a shared edge (see Figure 4.25).
The test environment thus contains no transitions in visibility but does have a
region (along the shared edge) at which the kernel of the area-area integrand
becomes singular.
Results are shown in Figure 4.26. Although in all the algorithms the mean
tends toward an accurate answer, any single invocation of the Monte Carlo
method can return an extremely inaccurate answer due to the singularity. The
standard deviation is much less in the hemicube and Malleys method since the
area-hemisphere method removes the singularity analytically before the numer-
ical evaluation.
There are a number of other ways to deal with the singularities. If the
singularities in the kernel are guaranteed to occur only on the edges of the
elements (e.g., where a wall meets the floor), and the basis set is chosen to have
zeros of multiplicity 2 at its edges, then the singularity will disappear when
weighted by the basis function under the integral. One example of such a basis
is provided by the Jacobi polynomial set on the element parameterized by u and
v ranging from 1 to 1:

N(u, v) = (1 u2)2 (1 v2)2 (4.36)

Although the numerator goes to zero as fast as the denominator, this of course
does not prevent a divide by zero in practice. One would need explicitly to
check for this condition. This option is not pursued further here. Details of this
approach can be found in [270].
Another solution is to move the quadrature point away from the singularity
by a small amount. In other words, if r is below some threshold, the quadrature
points are moved away from the intersection of the elements where the singu-
larity occurs. Care must be taken that the movement is inward on the elements.
Since there is no analytic justification for moving the points inward, the accu-
racy of the results will not converge to zero with large numbers of samples, but
the possible large errors incurred from hitting the singularity will be avoided.
Hybrid methods that handle singular regions analytically coupled with numer-
ical methods for the rest of the integral are also possible but have not been
extensively explored.
In point collocation, nodes on concave corners, such as on the floor where
it intersects the wall, will result in a form factor of zero for the intersecting
surface. In the case of a projection algorithm like the hemicube, this is because
the intersecting surface passes through the eye point and thus projects as a line
[163]. The form factor will be zero for other algorithms, since the cosine of the
angle with respect to the intersecting surface is zero. If that surface reflects a
significant amount of energy, a visible dark streak along the receiving surface
near the intersection may result. Moving the node inward may help in these

Radiosity and Realistic Image Synthesis 101

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.13. NONCONSTANT BASIS FUNCTIONS

Figure 4.26: Test results on perpendicular elements.

Radiosity and Realistic Image Synthesis 102

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.14. ACCELERATION TECHNIQUES

cases. The hierarchical subdivision algorithms discussed in Chapter 7 provide a

more systematic approach to this problem.

4.14 Acceleration Techniques

The form factor computation is expensive and is repeated many times during the
radiosity solution. It has thus been a natural target for acceleration. Improved
algorithms and hardware assistance have both been applied, particularly to the
visibility problem. In the following sections we will survey methods for speeding
up the hemicube and ray-tracing-based form factor algorithms. (A discussion of
coarser grained parallel algorithms will be held off until Chapter 5.)

4.14.1 Hemicube Acceleration

Visibility Preprocessing

Performing a single hemicube consists essentially of determining the surfaces

visible from a particular point of view. During the radiosity solution this visi-
bility calculation is repeated over and over for potentially thousands of different
viewpoints. Since the geometry of the scene does not change during this process,
radiosity is a natural candidate for visibility preprocessing.
In a visibility preprocess, the surfaces of the scene are organized by a spa-
tial data structure that allows the visibility from any particular viewpoint to be
quickly determined. Visibility preprocessing has a long history in computer
graphics; the need to render a fixed geometry from many different viewpoints is
often encountered in interactive simulation or walk-through and fly-through
applications. Algorithms date back to Schumakers list priority method [208]
and include octree [72] and BSP-based [87] visibility algorithms.
Wang and Davis [249] build an octree as a visibility preprocess for radiosity.
As space is split recursively into octants, surfaces are subdivided into elements by
clipping them against the octant boundaries. As a result, when the preprocess
is complete each element will fall into one and only one octree cell. For a
particular hemicube, the goal is to traverse the octree cells in front-to-back
order with respect to the hemicube location. The elements can then be painted
onto the hemicube faces in front-to-back order, which eliminates the need to
compute and store the z-depth.
For a particular viewpoint, a visibility priority can be established for the
eight children of a given octant. The subdivision of the octant determines eight
subspaces. The priority ordering of the eight children is determined by which of
these subspaces the viewpoint falls into. Each of the children is visited according
to the priority ordering and traversed recursively in the same fashion. The leaves
of the octree are thus visited in front-to-back order with respect to the viewpoint.

Radiosity and Realistic Image Synthesis 103

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.14. ACCELERATION TECHNIQUES

Figure 4.27: Hemicube positioning and clipping.

If more than one element is contained in a single leaf cell, the priority of these
must be established by some other means.

View Coherence

Although a hemicube requires projecting the scene onto five different image
planes, the five planes are obviously far from arbitrarily related. The hemicube
determines five orthogonal 90-degree view frustums. If the five planes are posi-
tioned to face in the +X, X, + Y, Y, and +Z directions (see Figure 4.27),
the coherence that arises from this arrangement provides several opportunities
for optimization.
Vilaplana and Pueyo [245] observe that once the elements have been trans-
formed to the view coordinates of the front face of the hemicube, they can
be rapidly transformed for the other faces simply by swapping coordinates and
changing signs appropriately. Given a coordinate x, y, z in the top faces view,

Radiosity and Realistic Image Synthesis 104

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.14. ACCELERATION TECHNIQUES

the transformations are

Left : (x, y, z) = (y, z, x)

Right : (x, y, z) = (y, z, x)
Front : (x, y, z) = (x, z, y) (4.37)
Back : (x, y, z) = (x, z, y)

Vilaplana and Pueyo also suggest a way of accelerating the clipping that
must be performed against each of the five view frustums. The top view frustum
shares a clipping plane with each of the four side view frustums, and each of the
side frustums shares planes with two other side frustums. Vilaplana and Pueyo
describe a modified Sutherland-Hodgman clipping algorithm in which each edge
of each element is clipped against all five frustums in one pass (as well as against
the sixth horizon plane determined by the base of the hemicube). Vertices are
transformed into the view coordinates for the top face. Each vertex of an element
is then classified in turn by testing against the six clipping planes. The vertices,
including those created by clipping edges, are added to lists maintained for each
frustum. Following clipping, the vertices for each frustum are transformed as
above, projected onto the corresponding hemicube face, and scan-converted as
usual.

Hardware Z-Buffering
One inspiration for the original hemicube algorithm was the availability of hard-
ware depth-buffers to do the visibility determination. The basic approach to
using graphics acceleration is to specify a view for each face of the hemicube
and pass the elements down the hardware pipeline for scan-conversion and Z-
buffering. Instead of a color, an item ID is stored at each pixel, which identifies
the visible element. The ID can be generated by assigning each element a color
corresponding to its ID. The colors can then be read back from the frame buffer
and decoded, with the contribution due to the delta form factor at each pixel
computed in software as usual. This approach has been implemented by Baum
et al. [21] and others. Bu [39] also describes a hardware implementation for fast
form factor computation.
A similar approach is used by Recker et al. [191]. Instead of the hemicube,
Recker uses a single plane method, which requires that each element be pro-
jected, clipped and scan converted once rather than five times.
An interesting hardware alternative to the hemicube has been proposed by
Fuchs et al. [88] for the Pixel Planes 5 graphics engine. This algorithm computes
Nusselts analog directly, taking advantage of the ability of Pixel Planes 5 to
evaluate a quadratic expression at every pixel. Their estimated time to compute
form factors to 100,000 elements is 1 second.

Radiosity and Realistic Image Synthesis 105

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.14. ACCELERATION TECHNIQUES

4.14.2 Ray Tracing Acceleration

One advantage of ray tracing for computing form factors is the wealth of acceler-
ation algorithms and data structures that are available. Bounding box hierarchies,
voxel grids, and octrees are all applicable to ray tracing for radiosity. 14
The rays used for form factor computations are generally shadow rays. Thus
they need only determine whether occlusion occurs or not, as opposed to de-
termining the closest intersection alone the ray. There are several specialized
tricks for accelerating shadow ray testing that are worthwhile looking at for ra-
diosity implementations [111]. Perhaps the simplest, shadow caching, is always
to test the ray against the most recently encountered occluding object, on the as-
sumption that an extended object will tend to intersect a number of neighboring
rays.
Directional acceleration schemes are particularly well suited to radiosity.
These take advantage of the coherence between rays leaving a particular region
of space in a particular set of closely related directions [11]. As has been seen,
computing the form factor from one element to another may require shooting
many rays between the two elements. These rays tend to have similar origins
and directions.
The shaft-culling algorithm described by Haines and Wallace [112] and by
Zhang [271] is intended to exploit this coherence. Haines and Wallace organizes
the objects of the scene into a bounding box hierarchy. Rays can always be
tested for an intersection by traversing this hierarchy starting at the root node.
Intuitively, however, if many rays are to be shot between two limited regions,
such as two elements, it is worth some effort to create a short candidate list of
objects for testing. Candidate objects are those that lie in the region of space
joining the starting and ending areas.
Haines and Wallace first construct a volume, or shaft, using planes to join the
bounding boxes of the objects or elements that define the endpoints of the rays.
The bounding boxes are constrained to be orthogonal to the world coordinate
axes. This allows a number of geometric and algebraic simplifications in the
construction of the shaft. An example is shown in Figure 4.28.
Once the shaft is created, it is tested against the bounding box hierarchy.
A box that is entirely contained within the shaft is added to the candidate list.
A box that partly intersects the shaft is usually opened up, and its sons tested
against the shaft recursively, in a depth-first manner. In some cases it may be
better to add the box on the candidate list without opening it up. A box that
is entirely outside the shaft is culled, and traversal of that branch of the tree is
ended. The final result is a candidate list of bounding boxes and objects.
14
The best introduction to this topic is Arvo and Kirks Survey of Ray Tracing Accel-
eration Technique in [97].

Radiosity and Realistic Image Synthesis 106

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 4. THE FORM FACTOR
4.14. ACCELERATION TECHNIQUES

Figure 4.28: Shaft culling: first an axis aligned bounding box in constructed
around each polygon. Then a convex hull (or shaft) is constructed between pairs
of bounding boxes.

Shaft culling is potentially useful for occlusion algorithms other than ray
tracing. For example, the polygon contour clipping algorithm of Nishita and
Nakamae [175] could be accelerated using shaft culling to produce a candidate
list of polygons for clipping.
Marks et al. [160] describe an algorithm related to shaft culling in which the
object is to rapidly cull away empty portions of the volume connecting two
elements. The volume is subdivided recursively, with recursion stopping
when a volume becomes empty. After a certain subdivision limit is exceeded, the
remaining volumes are tested for occlusion using rays.
Teller and Hanrahan investigate incremental visibility algorithms that rapidly
preprocess the environment and categorize pairs of surfaces as fully visible to one
another, fully blocked, or partially blocked [232]. They also maintain a candidate
list of blocking surfaces between each pair. This preprocessed information then
greatly accelerates visibility calculations during form factor computation.

Radiosity and Realistic Image Synthesis 107

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS

Chapter 5

Radiosity Matrix Solutions

In Chapter 3, techniques were described for approximating the radiosity integral
equation (3.1) by a system of linear equations. Depending on whether point
collocation or Galerkin techniques are used, a system of linear equations is ob-
tained (equations 3.13 or 3.20) which when solved provide the nodal values that
determine an approximate radiosity function. In either case the linear equations
can be summarized by
KB = E (5.1)
where K is a matrix of interactions, B is a vector of nodal radiosity values, and
E is a vector of emission terms. The K matrix can be broken down into the
difference of a (almost) diagonal matrix and a matrix of coefficients scaled by
reflectivity terms, l r:
[MPF] B = E (5.2)
If the radiosity function is approximated using constant basis functions, area
terms can be divided out, and M is simply the identity I. Thus K looks like

1 1F1,1 1F1,2 1F1,3 . . 1F1,n

F 1 2 F2,2 2 F2,3 . . 2 F2,n
2 2,1

. . (5.3)

. .
n1Fn1,1 .

n Fn,1 . . . . 1 n Fn,n

The previous chapter concentrated on ways to evaluate the entries of F, the

form factors. This chapter focuses on how to solve the linear system given K
1
As noted in Chapter 3, the reflectivity, r, must be specified for each of the wave-
lengths or color bands of interest. In most applications for image synthesis, 3 or 4
wavelengths are sufficient. Thus, there will be 3 or 4 K matrices, one for each wave-
length. The form factors discussed in Chapter 4 are only dependent on geometry and
are thus valid across the visible spectrum. The radiosity problem will continue to be
addressed as a monochrome one, with the understanding that the algorithms described
are applicable at each wavelength individually.

Radiosity and Realistic Image Synthesis 109

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.1. QUALITIES OF THE MATRIX

and E. The characteristics of K are examined first followed by a description

of applicable techniques from the numerical methods literature on solutions of
linear systems. Finally these methods are applied to the radiosity problem.

5.1 Qualities of the Matrix

The following points describe qualities of the matrix K that are relevant to the
selection of specific numerical algorithms:

Size: In general, the matrix will be square and of size n by n, where n is

the number of basis functions (i.e., the unknown coefficients Bi) that make
up the approximation of the radiosity function. In Chapter 7 hierarchical
methods are examined in which rectangular matrices will appear.
Sparcity: As for most systems resulting from integral equations of a form
like the radiosity equation, the matrix will typically be more full than
sparse. The ijth position of the matrix will be zero only when the reflec-
tivity of a surface is zero (i.e., the surface is black) or the supports of the
ith and jth bases are completely invisible to one another (e.g., for constant
bases, when element i cannot see element j). The bases can be invisible
to one another either because of occlusion or because they face away from
one other. Only an environment resembling the inside of an empty sphere
(i.e., one in which every element can see every other element) will result
in a completely full matrix. For a complex environment, the matrix will
be fairly sparse, since many elements will be invisible to each other. In
fact, if the environment consists of two separate rooms with no means for
light to travel between them, the matrix will be block diagonal, meaning
that, intuitively, it can be divided into two independent subproblems.
Symmetry: A matrix A is symmetric if elements ai j = aj i for all i, j. K
is not symmetric in its current form. However a simple transformation can
produce an equivalent system in which the matrix is symmetric. If each
row i is multiplied by the area of the ith element, then the F matrix is made
symmetric, due to the reciprocity of form factors (i.e., FijAi = FijAi). K
can also be premultiplied by the inverse of the reflectivities to complete the
symmetry, but zero-valued reflectivities (as in the case of light sources)
will be a problem. Thus, although there is a type of symmetry in the
physical process which is manifest in the matrix, it will not be relied on
explicitly in the solution methods discussed in this chapter.
Diagonal Dominance: A matrix is said to be diagonally dominant if the
absolute value of the sum of the off-diagonal terms in each row is less

Radiosity and Realistic Image Synthesis 110

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.1. QUALITIES OF THE MATRIX

than or equal to the absolute value of the diagonal term itself:

|K
j =1
ij | | Kii |, i (5.4)
j i

This is strictly true for the radiosity problem with constant basis functions.
This can be seen by examining the matrix equation 5.3, in which the
diagonal terms are all equal to one. The sum of the form factors in any
row is by definition equal to unity, and in any physically realistic system
the reflectivity terms will be less than one. Thus, the off-diagonal terms
must sum to less than the diagonal term. The matrix will also be diagonally
dominant for general basis functions with local support.
The diagonal dominance of K ensures that particular iterative solution
methods such as Gauss-Seidel (described in section 5.3.2) will converge.
However, the diagonal dominance of rows should not be confused with the
diagonal dominance of columns, which does not characterize the current
form of K. This distinction will arise in the discussion of Southwells
method, described in section 5.3.3.
Spectral Radius: The spectral radius of a matrix is a particular type of
norm that describes the size of its largest eigenvalue. The detailed defi-
nition of these terms is beyond the scope of this text. The spectral radius
is an indicator of the speed with which iterative methods will converge.
Intuitively, one can see that as the reflectivity values approach unity (i.e.,
perfect reflectors), the system will require more iterations to converge.
Physically speaking the light will reflect about the environment more be-
fore being absorbed. Mathematically, the spectral radius is approaching
one.
If PF has a norm less than one, then K = [ I PF] is invertible, and the
Neumann series of successive multiplications of PF will converge to the
inverse.

If PF < 1 then K -1
= [ I PF ]
-1
= (PF)
a= 0
a
(5.5)

This property is also important for iterative methods such as Jacobi itera-
tion, which in essence mimics the Neumann series.
Condition: The condition number of a matrix describes how sensitive the
solution is to small perturbations in the input (as in the case the emission
terms). In general, the matrix arising in radiosity applications is well
conditioned, indicating that most solution methods are applicable.

Radiosity and Realistic Image Synthesis 111

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.2. LINEAR SYSTEM SOLUTION METHODS

5.2 Linear System Solution Methods

Of the great variety of algorithms that have been developed to solve linear
systems, the following sections will concentrate on methods that have proven
useful for the radiosity problem. Each of the basic algorithms will be described,
followed by a discussion of its behavior in the context of the radiosity problem.
More general surveys and more theoretical treatments are widely available in
numerical methods texts [138, 185, 226].
5.2.1 Direct Methods
Direct methods for solving linear systems are perhaps most familiar but are
not well suited to large systems of equations. Such methods like Gaussian
elimination and its variants, can be applied to systems such as the radiosity
problem, but they exhibit a computational complexity related to the cube of the
number of equations, O(n3). Thus, these methods are prohibitively expensive
except when applied to small problems, or when the system of equations is
sparse. For image synthesis the system of equations is liable to be quite large
and relatively full. Thus, iterative solution methods are the focus of the ensuing
discussion.
5.2.2 Iterative Methods
In contrast to direct methods, iterative methods begin with a guess for the solu-
tion and proceed by performing (preferably) inexpensive operations that move
the guess to a better guess. The solution is said to have been found, or con-
verged when there is some confidence that the current guess is very close to the
actual solution. Given the linear system

KB = E (5.6)

containing the unknown vector, B, and an initial guess, B(0), the error, e(0)
defined to be equal to the difference between the actual answer B and the current
guess

e(0) = B B(0) (5.7)

Since the real answer B is not known, e cannot be measured directly. However,
one can define a residual, r(0), where

r(0) = KB(0) E (5.8)

Clearly, if the residual is zero, then the solution guess is correct and the error
is zero as well. In contrast to the error, the residual is a quantity that can be
directly measured.

Radiosity and Realistic Image Synthesis 112

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

After each iteration2 of the solution process, the most recent guess and
residual, (B(k) and r(k)), are replaced by a more accurate guess and residual,
(B( k+1) and r( k+1)).
The initial guess, B(0), can influence the rate of convergence. In general,
an initial guess that is closer to the final solution will require fewer iterations.
If there is no information available upon which to base an initial guess, B(0)
can simply be a vector of zeros. For the radiosity problem the light source
radiosities are given a priori by E, and a better initial guess is B(0) = E since
light sources typically have zero reflection.

5.3 Relaxation Methods

The types of iterative algorithms explored below are called relaxation methods.
The idea is that at each step of the algorithm one element of the residual vector
will be set to zero. In other words, one of the B(k) will be chosen to change
in such a way that ri(k+l) = 0. Of course the other rj(k), j i may increase,
but hopefully an improvement has been made on the whole. The input to each
step may be the approximate solution B(k) (and perhaps part of the new guess
B(k+l)) and one or all of the current residuals, in addition to the matrix K and
vector E.

5.3.1 Jacobi Iteration

Perhaps the simplest iterative scheme is to update each element Bi(k) of the
solution vector by solving for that variable using the current guess Bi(k). Thus
n steps (one iteration) can be performed simultaneously (one for each i). Solving
for a single Bi, beginning with the ith row from the matrix equation 5.3

K Bj = E
j
ij i (5.9)

moving all but the ith term to the right hand side

Kii Bi = Ei K B
j i
ij j (5.10)

2
Superscripts (e.g., B(k)) will be used to indicate a complete iteration of the iterative
solution process. A superscript of zero indicates the state prior to the first iteration. An
iteration is distinguished from a step, which refers to the processing of one entry in B
during a particular iteration. A complete iteration will usually involve n steps. Subscripts
will be used to indicate position in the vector when necessary.

Radiosity and Realistic Image Synthesis 113

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

and dividing by Kii results in a new value for Bi for iteration k + 1

B(k )
Bi(k +1) = Ei K
j i
ij
j
Kii
(5.11)

With the definition of the residual vector (equation 5.8), this can be simplified
to adding the residual divided by the diagonal to the previous guess,

ri(k )
Bi(k +1) = Bi(k ) + (5.12)
Kii

Thus if the residual vector is known3 the full iteration (all n steps) can be
performed in linear time, O(n). Although each step in isolation would relax one
element of the residual if the old solution were still valid, none other than the first
element of B dealt with would actually result in zeroing out the corresponding
residual. Thus this is not, strictly speaking, a relaxation method. As with all
new guesses. It is this recursive nature of iterative methods that makes them
very sensitive to the characteristics of the system outlined above.
A close look at the Jacobi algorithm reveals its essential similarity to the
Neumann series described in section 5.1. Thus, it can be said with confidence
that this simple algorithm will converge to the correct solution in the case of the
radiosity problem.
5.3.2 Gauss-Seidel Iteration
Gauss-Seidel iteration is a slight variation of the Jacobi method. It provides
a true relaxation method and will usually improve the speed with which the
solution converges. At each step during an iteration, the Gauss-Seidel method
uses the most up-to-date version of the solution vector, rather than the solution
vector computed by the previous iteration. Thus, at each of the n steps of an
iteration, the new entry of B is computed using the values B(k+1) computed by
previous steps of the current iteration. Otherwise, values computed during the
previous iteration B(k) are used. Thus, to relax ri, set

i 1 +1) n
B(k B(k )
Bi(k +1) = Ei K
j =1
ij
j
Kii
K
j =i +1
ij
j
Kii
(5.13)

3
Although Jacobi iteration is not often used on its own since more efficient methods
are known, it will become clear in later sections how one can take advantage of the
known residual vector to increase the power of other algorithms.

Radiosity and Realistic Image Synthesis 114

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

1 for ( all i ) Bi = starting guess ;

2 while ( not converged ) {
3 for ( each i )
4 Bi = Ei nj =1, j i BKi Kiiij ;
5 }
6 output B ;

Figure 5.1: Pseudocode for Gauss-Seidel iterative solver.

Performing the step for a single entry of B requires taking the dot product
of the current guess, B(k),(k+1), with a row of the matrix, and thus requires
O(n) operations. Consequently, a full iteration consisting of n steps is O(n2).
Alternatively, since
n
ri(k ) = Ei K B
j =0
ij
(k )
j (5. 14)

if all the residuals are updated after each step, one can set
ri(k )
Bi(k +1) = Bi(k ) + (5.15)
Kii
In this case the cost of a step is reduced to O(1), but the residual update will
be O(n) for each step and a full iteration is again O(n2). Essentially, these
operations are equivalent to the operations in equation 5.12.

The Gauss-Seidel Algorithm

If the i s are relaxed in order, the Gauss-Seidel algorithm in Figure 5.1

results. One iteration of the algorithm involves relaxing each residual in turn.
Gauss-Seidel iteration can be shown to be absolutely convergent for diagonally
dominant systems such as the one arising in the constant radiosity formulation. A
variant of this algorithm has been shown to work well for the radiosity problem
[62]. Convergence criteria are typically specified by a maximum residual, || r|| `,
or a maximum change in the solution, ||Bk+1 Bk || ` , where the notation
||f (x)|| ` refers to the l ` norm (see box page 133).

Gathering: A Physical Interpretation of Gauss-Seidel

Examining a single step in the Gauss-Seidel algorithm (line 4), a single nodal
radiosity value, Bi, has been selected to update by summing contributions from

Radiosity and Realistic Image Synthesis 115

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

Figure 5.2: Gathering: a Gauss-Seidel step.

all other radiosity values weighted by the reflectivities and form factors in the
ith row of K. Assuming constant elements, each term of summation in line 5
can be expanded using the K matrix (equation 5.3) entries and taking advantage
of the fact that Kii = 1 if elements cannot see themselves. Thus

DBi = riBiFij (5.16)

This term represents the contribution made by element j to the radiosity of

element i. Thus the complete step is equivalent to gathering the light from all
other elements4 to arrive at a new estimate for Bi (see Figure 5.2). Note that
single step involves one row of K and updates a single radiosity value.

5.3.3 Southwell Iteration

A variation of the Gauss-Seidel method provides an alternative algorithm with
a slightly different physical interpretation. In the less well known Southwell
iteration [90], rather than relax each residual in turn, the row i with the largest
residual, Max (r), will always be selected.
4
Strictly speaking, gathering is from the energy under the other basis functions, but
for constant bases one can speak of gathering from elements.

Radiosity and Realistic Image Synthesis 116

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

Thus a single step,5 p, of the Southwell algorithm is

B j Kij
For i, such that ri = Max(r) : Bi = Ei (5.17)
j i
Kii

It would seem that Southwell would require O(n2) operations to compute

all the ris before picking the greatest one. (The computation of each ri above
involves computing the dot product of B(p) with the row Ki.) Fortunately, if
at some step p, r(p) is known for a given B(p), the next approximation can be
made if the changes B(p) are known. Thus,
B(p+1) = B(p) + B(p) (5.18)
and the updated residual can be computed as:
r(p+1) = E K(B(p) + B(p)) = r(p) KB(p) (5.19)
since
r(p) = E KB(p) (5.20)
However, in this case all the B are zero except Bi. Therefore,
( P)

K ji
r (j p+1) = r (j p) ri( p) , j (5.21)
Kii

and expanding K for the case of constant elements,

rj(p+1) = rj(p) + rjFji * ri (p), ;j (5.22)
Updating r thus takes only O(n) steps involving multiplying a scalar by a column
of the matrix, K.
The final requirement is to compute r(0) easily at the start of the algorithm.
Simply choosing B(0) to be 0 (the zero vector) gives
r(0) = E KB(0) = E (5.23)
The Southwell Algorithm
The Southwell relaxation method follows the pseudocode in Figure 5.3.
As in the Gauss-Seidel algorithm, each step is O(n). It is more difficult to
characterize the cost of an entire iteration since some elements may be revisited
many times before another is operated on. As will be seen below, a close
approximation to the final answer can often be achieved in a small number of
steps, resulting in a linear overall solution time.
5
It is impossible to write down a single iteration, since it is possible that one residual
will be relaxed many times before another is ever touched. The superscript p is used
to indicate the step number in this section rather than the iteration number as in other
sections

Radiosity and Realistic Image Synthesis 117

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

1 for (all i ) {
2 Bi = 0 ;
3 ri = Ei ;
4 }
5 while ( not converged ) {
6 pick i, such that |ri| is largest ;
7 Bi = Bi + ri / Kii ;
8 temp = ri ;
9 for ( all j ) rj = rj Kji/Kii * temp ;
10 }
11 output B ;

Figure 5.3: Pseudocode for Southwell relaxation.

Traditional Radiosity Progressive Refinement

Mesh Environment Mesh Environment

Compute Form Factor Matrix Compute One Row of F

Solve Linear System Perform One Step of Solution

Display Result Display Intermediate Result

Figure 5.4: Traditional radiosity versus progressive refinement.

Shooting: A Physical Interpretation of Southwell Iteration

Each step of the Southwell algorithm does two things. It updates one element
of the radiosity vector, Bi, and updates all of the residuals. If constant bases
are used, the physical interpretation of B is the vector of nodal radiosity values
and thus the radiosity of each element. What is the physical interpretation of the
residuals, r? The residuals are the unshot radiosity of each element. In other
words, the residuals represent the radiosity that has been received as illumination
by the elements but has not yet been reflected or shot back out to contribute
further to the illumination of the environment.
It is clear from equation 5.22 that each step takes one residual and adds
fractions of it to each of the others. The specific fractions are the ith column of

Radiosity and Realistic Image Synthesis 118

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

Figure 5.5: Shooting: a progressive radiosity step.

K. Rewriting equation 5.22 using the reciprocity principle,

Ai
r (j p+1) = r (j p) + r j F ji ri( p) = r (j p) + r j Fij ri( p) (5.24)
Aj

or rearranging,
rj(p+1) Aj = rj(p) Aj + rj Fij ri(p) Ai for each residual rj (5.25)

shows that the total unshot energy (radiosity 3 area), is reduced in each step
since both the r and the sum of the Fijs are less than one.
Progressive Refinement
Cohen et al. [59] describe a slightly different form of Southwells algorithm,
called progressive refinement or progressive radiosity. The goal of progressive
refinement is to display results after each step of an iterative process, in order
to provide immediate feedback. The requirement is thus not only to provide an
algorithm that converges quickly, but one that makes as much progress at the
beginning of the algorithm as possible. In progressive refinement the flow of
the traditional radiosity approach becomes an iterative repetition of the steps of

Radiosity and Realistic Image Synthesis 119

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

1 for ( all i ) {
2 Bi = Ei ;
3 Bi = Ei ;
4 }
5 while ( not converged ) {
6 pick i, such that Bi * Ai is largest ;
7 for ( every element j ) {
8 rad = Bi * rjFji ;
9 Bj = Bj + rad ;
10 Bj = Bj + rad ;
11 }
12 Bi = 0 ;
13 display the image using Bi as the intensity of element i ;
14 }

Figure 5.6: Pseudocode for progressive refinement.

computing form factors, performing steps of the solution and displaying inter-
mediate results (see Figure 5.4). For constant elements, the progressive radiosity
algorithm follows the pseudocode in Figure 5.6.
The previous algorithm has the following physical interpretation. All el-
ements i have a value Bi, which is the radiosity calculated so far for that
element, and Bi which is the portion of that elements radiosity that has yet
to be shot. During one iteration, the element with the greatest unshot radiosity
is chosen and its radiosity is shot (see Figure 5.5) through the environment. As
result of the shot, the other elements, j, may receive some new radiosity,
rad. This rad is added to Bj. rad is also added to Bj since this newly
received radiosity is unshot. As a result of the shooting, element i has no unshot
radiosity, so Bi = 0.
In this algorithm one shooting step (lines 711 in Figure 5.6) updates all
the other elements. Energy is shot from the element that currently has the most
unshot radiosity. One shooting step takes O(n) operations and can be viewed
multiplying the scalar Bi by a column of the form factor matrix. Cohen et
al. [59] show that in many cases only a small fraction of n shooting steps is
required to approximate a solution closely.
At first glance Southwell and progressive refinement seem to be quite dif-
ferent. Southwell updates only one entry in B, while progressive refinement
updates them all at each step. However, recognizing B in the progressive refine-
ment algorithm to be the sum of B and r of the Southwell algorithm, the two

Radiosity and Realistic Image Synthesis 120

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

Gauss-Seidel after 1, 2, 24, and 100 Steps

Progressive Refinement after 1, 2, 24, and 100 Steps

Figure 5.7: Guass-Seidel versus progressive radiosity.

Radiosity and Realistic Image Synthesis 121

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

algorithms are in fact almost identical. There is one other difference: progres-
sive refinement selects the largest energy to shoot (i.e., the largest ri Ai) rather
than simply the largest residual. The use of the sum of the already shot radiosity
and the unshot residual to be the current answer can be explained as a hybrid of
Southwell and Jacobi iteration. This sum is exactly equivalent to the Southwell
algorithm followed by one full iteration of Jacobi iteration before display.
It should be clear why this algorithm should converge faster than a Gauss-
Seidel algorithm, particularly in the early steps (see Figure 5.7). The Southwell
and progressive refinement algorithms make very good guesses about which el-
ement of the residual vector to relax to get the greatest reduction in the residual
as a whole. Particularly at the start of the iterations, all the unshot energy resides
in the light sources (i.e., only a few emission terms are nonzero) thus there a
only a few nonzero residuals. As the light interreflects about an environment,
the unshot energy (i.e., residuals) becomes more evenly spread out and thus the
advantage of selecting the largest to process is reduced as the algorithm pro-
gress. A full analysis and proof of convergence properties of these algorithms
for the radiosity problem can be found in Gortler et al. [101].

5.3.4 Ambient Energy and Overrelaxation

Ambient Term
In Cohen et al. [59] an additional term, similar to the ambient illumination
term commonly used in ray tracing, is added to the radiosity approximation,
for the purpose of display only. The ambient term is a crude approximation
that accounts for the reflected illumination not yet accounted for in the solution.
The ambient term models the unknown illumination arriving from unknown
directions as a constant illumination from all directions.
At each step of the progressive refinement solution, there is a known amount
of energy that has not yet been distributed or shot. The average unshot radiosity,
B , is simply the sum of the entries in the residual vector weighted by the
fraction of the total area6 of the element corresponding to each entry.

ri Ai
B = (5.26)
Ai

Each time some of this unshot radiosity is processed in a step of the progressive
refinement algorithm, some of that energy is absorbed, but some of it is returned
to the unshot quantities of other elements (or basis functions). On average, with-
out knowing in advance where the unshot energy will arrive, one can estimate
6
In the case of nonconstant basis functions, the area is the integral of the basis function
dotted with itself.

Radiosity and Realistic Image Synthesis 122

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

Displayed Image after 1, 2, 24, and 100 Steps

Figure 5.8: Progressive radiosity with ambient addition.

that some fraction of the energy will be reflected. This fraction is represented
by r ,where
r i Ai
r= (5.27)
Ai

Of course, in the same way that some of the shot energy is reflected, some of
the reflected energy will be rereflected and so on. The total reflection, Rtotal,
can be represented by the infinite sum
1
Rtotal = 1 + r + r 2 + r 3 + .... = (5.28)
1 r
Finally, the product of the average unshot radiosity, B , and the total reflection
gives us an estimate for the ambient radiosity, B ambient:

Bambient = BRtotal (5.29)

Each element i will reflect its own fraction ri of this quantity. Thus for display
purposes only (see Figure 5.8)

Bidisplay = Bi + ri Bambient (5.30)

Radiosity and Realistic Image Synthesis 123

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

Figure 5.9: Convergence versus number of steps for three algorithms.

Adding the ambient term brings the displayed image much closer to the
converged solution very early in the process. Figure 5.9 shows a plot of the
convergence for the images in Figures 5.7 and 5.8 of the Gauss-Seidel, progres-
sive refinement, and progressive refinement plus ambient versus number of steps
(not iterations). The vertical axis is in terms of area weighted RMS error:

p 2
(Bi Bi ) Ai
100 (5.31)
Ai

where p is the step number, and B* is the result of a converged solution.

It should be noted that as each step of the progressive refinement algorithm
proceeds, the total unshot energy is reduced and thus the ambient term is also
reduced and the displayed and actual radiosity values, converge.
Overrelaxation
The ambient term described in the previous section is a crude representation
of the energy that has not yet been shot. It improves the displayed image, but
does not speed up convergence of the actual solution. Overrelaxation techniques

Radiosity and Realistic Image Synthesis 124

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.3. RELAXATION METHODS

provide a more systematic attempt to acknowledge the energy unaccounted for

so far, by pushing the solution farther ahead at each step than the current
solution vector indicates.
Overrelaxation techniques are similar to the relaxation methods previously
described with one exception. When relaxing a particular residual, the change
in the solution vector is increased by a factor v (or equivalently the residual is
considered to have been larger than it actually is). Equation 5.12 becomes

ri(k )
Bi(k +1) = Bi(k ) + v (5.32)
Kii

and the ith residual being relaxed is now not set to zero, but rather,

ri(k+1) = (1 v) ri(k) (5.33)

Overrelaxation involves a value for v greater than 1, while underrelaxation

involves a value between 0 and 1. In cases where Gauss-Seidel methods con-
verge, overrelaxation will often increase the convergence rate. (Underrelaxation
is sometimes useful for unstable systems.) The best value for v will depend on
the behavior of the particular problem being solved and is usually determined
experimentally. For radiosity algorithms, overrelaxation with v equal to about
1.2 has been found to give good results [58].
In the context of the progressive refinement method, overrelaxation can be
thought of as overshooting. Feda [82] discusses a heuristic for overshooting in
the progressive refinement algorithm which works well if certain restrictions are
placed on the overshot amount to avoid divergence.
Gortler et al. [101] propose an analytic approach that simulates the known
interreflection between the shooting element and all others while maintaining a
linear complexity at each step. In each step all the selected elements residual
is relaxed as well as the portion of the residual from all other elements to the
selected element. The method is called a super-shoot-gather, since in essence
light is shot from the element to all other elements, and then gathered back
to the shooting element, thus taking further advantage of the form factors that
have been computed for the shooting step. This form of overrelaxation was
found to be particularly useful in bright environments (i.e., environments with
a high average reflectivity). This should be no surprise since it is the secondary,
tertiary, etc., reflections that are captured by the overshooting.

Other Solution Methods

The above sections do not completely enumerate all methods for solving linear
systems. Methods applicable to the radiosity problem have been addressed to

Radiosity and Realistic Image Synthesis 125

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.4. DYNAMIC ENVIRONMENTS

the exclusion of others. However, a class of applicable methods based on a

divide and conquer paradigm remains to be described. This approach will be
discussed in Chapter 7, where hierarchical solution methods are outlined.

5.4 Dynamic Environments

All the discussion of the radiosity method in previous chapters has assumed
that that environment is unchanging, that is, the geometry, lighting, and material
properties are static. Although this is acceptable for some applications, many
other applications of image synthesis require the ability to change or move
objects. Of course, the solution can always be restarted from scratch after
each change, but this requires throwing away potentially valuable information,
particularly if the change is small.

5.4.1 Lighting Changes

Changes in the emission terms for the light sources are the simplest type of
change to include. The form factors are purely a function of geometry, thus the
most computationally expensive part of the radiosity algorithm is not affected
by changes in the lighting. Thus the K matrix is unchanged and modifications
to E require at most a new solution to the linear system of equations, KB = E.
If the changes to E are small, then the old solution is a good approximation to
the new one and should be used as the starting guess for the iterative methods
discussed above. For example, one might begin with the old solution and shoot
out the additional light from a light source that has changed. This also includes
the possibility of shooting negative light from lights that have been turned down.
If the application calls for many lighting changes in an otherwise static
environment, a separate solution can be run for each light source with a unit
emission [5]. Since the radiosity equation is linear, the independent solutions
can then be scaled and summed to provide any possible setting for the lights
with no extra matrix solutions.

NumLights
B =
L =1
EL B L (5.34)

where EL= the emission of Lth light source, and BL = the radiosity solution
vector computed using the emission vector set to 1 for Lth light source and 0
elsewhere.
This provides a very rapid means of changing between lighting settings.
Dorsey [73] has gone one step further in her opera lighting design system by
prerendering multiple images of complex scenes from a fixed viewpoint, differ-
ing only by the light emissions (color plates 53, 54 and the back cover). By

Radiosity and Realistic Image Synthesis 126

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.4. DYNAMIC ENVIRONMENTS

Figure 5.10: A dynamic environment with bounding box around all positions of
moving object.

scaling and summing the images themselves, different lighting conditions can be
set at interactive rates. (More details of the rendering step and this application
can be found in Chapters 9 and 11). If the cost of inverting K is undertaken,
then the radiosities B due to any emission vector E can be solved by a simple
matrix-vector multiplication, K1 E = B.

5.4.2 Reflectivity Changes

Changes in the reflectivity of surfaces in the environment also do not require
new form factor computations. Again, at most this will require modifying the
rows of K based on the new r values and performing a new solution of the linear
equations. However, a new solution accounting for changes in a single, or a few,
surfaces can be found more quickly by starting from the original solution [51].
New unshot energy BjAj (possibly negative) can be shot from the surfaces
that have changed, where

r new
j r old
j
B j = Bj (5.35)
r old
j

Clearly, this will not work in the case of surfaces that were originally defined
as black (i.e., pjold = 0).

5.4.3 Changes in Geometry

Changes in the geometry present the most challenging problem for rapidly de-
termining new solutions. Any individual change in the shape or position of a

Radiosity and Realistic Image Synthesis 127

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.4. DYNAMIC ENVIRONMENTS

Figure 5.11: A dynamic environment: (a) initial solution, (b) after shooting
negative energy to table top, (c) after shooting positive energy with airplane in
new position.

single object, or the addition (or removal) of a new object, has the potential to
change any entry in K. However, if it is assumed that only small changes are
made, there are a number of acceleration methods that have been described and
demonstrated.
Baum et al. [20] describe a method that can be used when all changes in the
positions of any objects are known in advance of any form factor computation.
In this, case, a bounding object can be constructed for each dynamic object. The
bounding box contains the volume swept by the object as it moves along the path
(see the airplane in Figure 5.10). During form factor computation, form factors
are marked if their visibility is affected by the composite bounding box. Thus,
at subsequent points in time only marked form factors need to be recomputed
before a new solution is performed.
Chen [51] and George et al. [91] do not assume a priori knowledge of ob-

Radiosity and Realistic Image Synthesis 128

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.5. PARALLEL IMPLEMENTATIONS

Figure 5.12: A parallel radiosity implementation. One master processor controls

the overall flow of the computation and performs the row or columnwise vector
multiply for the solution. A display processor constantly updates the display
based on the partial solutions. The remaining processors (two are idle) compute
rows (columns) of form factors and report the results to the master processor. This
coarse grained approach assumes each processor has a full geometric description
of the environment.

ject motion, addition, or deletion. As in the methods for changes in lighting or

material reflectance, they start with an existing solution for an environment (see
Figure 5.11) and compute a new solution after some small change, for example
moving the position of a chair. The new solution proceeds by identifying the
affected form factors (e.g., between the light and the table top in Figure 5.11).
Negative energy is shot to undo the affected interactions. The dynamic object
is then moved to its new position, and positive energy is shot. The unbalanced
positive and negative energy are then propagated as before. Both Chen and
George et al. provide many implementation details concerning the rapid iden-
tification of affected form factors, the optimal ordering for the new shooting
sequences, and other issues.

5.5 Parallel Implementations

Radiosity solutions for complex environments continue to be slow. Beyond
improvements in the radiosity algorithms themselves, one is left with the possi-
bility of speeding up the solutions with the use of parallel hardware. A number
of implementations have been described that parallelize various portions of the
solution process.
The different implementations range from the use of built-in pipelined hard-
ware in graphics workstations [21], to course grained systems developed on

Radiosity and Realistic Image Synthesis 129

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 5. RADIOSITY MATRIX SOLUTIONS
5.5. PARALLEL IMPLEMENTATIONS

networked workstations [186, 191], to transputer arrays [46, 107, 188] to the
thousands of processors on SIMD machines such as the CM2 [77] and MasPar
[240]. Other reports describe theoretical analyses of possible algorithms [244].
Methods can be grouped according to where and how they exploit the natu-
ral parallelism in the algorithms. The computation of form factors is the primary
bottleneck in the radiosity algorithm, and it is on that step that most implemen-
tations have concentrated. In [46, 50, 81, 107, 190] several full rows of form
factors are computed in parallel. Each processor performs a single hemicube or
other form factor quadrature and reports the results to another processor which
uses the results for the next step in an iterative solution (see Figure 5.12). An
additional processor may be devoted to displaying the current state of the radios-
ity solution. In a finer grained approach, individual form factors are parceled
out to processors in [21, 95]. There are also many implementations of parallel
ray tracing which can be used at an even finer level within a single form factor
computation. Drucker and Schrder [77] exploit parallelism at many levels in
their implementation on the Connection Machine.
Each of the reports offers insights into the many subtleties that arise when
using and implementing parallel strategies. The reader is encouraged to seek
out these references for a fuller understanding of the algorithms and reports of
experiments on a variety of environments.

Radiosity and Realistic Image Synthesis 130

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING

Chapter 6

Domain Subdivision
As described in Chapter 3, the approximation of the radiosity function, B(x),
is a fundamental step in the discretization of the radiosity equation. The ap-
proximation, B (x), is defined by the linear combination of n basis functions,
Ni (x):

n
B(x) B(x)
= B N (x)
i=1
i i (6.1)

The basis functions define a finite subspace of functions from which the approx-
imation B (x) is realized by choosing the coefficients, Bi.
In a finite element approximation, each of the basis functions, Ni (x), has
local support; in other words, each is nonzero over a limited range of the function
domain. A basis function is associated with a single node and has a value of
zero outside of elements adjacent to the node.
As discussed in Chapter 3, each type of basis function is defined with respect
to a generic element, such as the unit square. The generic basis functions are
then mapped to the actual model geometry according to the subdivision of the
surfaces into elements. Thus, the form of any particular basis function is tied to
the placement and size of one or a few elements. The total set of basis functions
defining the approximation is thus determined by the mesh of elements and
nodes.
The next three chapters will describe strategies and algorithms for subdi-
viding the domain of the radiosity function into finite elements, with the goal
of producing an efficient and accurate radiosity solution. The accuracy of the
approximation, B (x), is influenced by the size, shape and orientation of the
elements, as well as by the polynomial order of the basis functions. An optimal
mesh uses as few elements as possible to achieve a desired accuracy, which
generally means distributing errors in the approximation as evenly as possible
among the elements.
To achieve a high quality mesh, it is first necessary to be able to measure the
accuracy of the approximation produced by a particular subdivision and choice

Radiosity and Realistic Image Synthesis 131

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.1 ERROR METRICS

of basis functions. It is also important to understand how various characteristics

of the mesh affect accuracy. This knowledge can then be used to develop
strategies for producing a good mesh. These basic issues will be discussed in
the present chapter. More sophisticated hierarchical mesh refinement methods
will be discussed in Chapter 7. The actual mechanics of subdividing geometry
will then be treated in Chapter 8.

6.1 Error Metrics

6.1.1 True Error

The true error in the approximate radiosity function at location x is the difference
between the actual function and the approximate function at x:

(x) = B(x) B (x) (6.5)

Of course, equation 6.5 cannot be evaluated directly, since B(x) is not known.
Instead, (x) must be estimated. Several error estimator are described in the
following sections. Particular implementations of these approaches for radiosity
will be described in section 6.3.3.

6.1.2 Local Estimate of Approximation Error

The mesh and associated basis functions determine a subspace of possible func-
tions. Higher-order basis functions increase the space of realizable functions,
allowing a closer approximation to the actual function. Thus, one approach to
estimating the error is to compare the original approximation
n

B(x) = B N (x)
i=1
i i (6.6)

where the basis functions, Ni(x), are of order k, to a higher order approximation
n

B(x) = B N (x)
i=1
i i (6.7)

where the new basis functions, N i (x), are of order k + 1 or greater. For example,
if B (x) is constructed using linear basis functions, the error can be estimated
by comparing B (x) to an approximation B (x) constructed using a quadratic or
cubic basis. The error estimate (x) is then given by

(x) = B (x) B (x) (6.8)

Radiosity and Realistic Image Synthesis 132

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.1 ERROR METRICS

Function Norms and Error Metrics

A function norm is a measure of the magnitude of a function, analogous
to a vector norm or an absolute value for a scalar. The Lp function norms
are given by
1/ p


p
Lp ( ) = = ( x) dx (6.2)
p

where is a function defined over domain .
When used to characterize the error of an approximation, (x) is simply the
difference between the desired and approximate solution: (x) = (x) =
f(x) f (x). The norm can be computed over a limited region to characterize
local error or over the entire domain of the function to measure global error.
The most commonly used norms are the L1, L2, or L norms, the cases
where p equals 1, 2 or for equation 6.2. The L1 norm of the approxi-
mation error, , is simply the area between the true function curve and the
approximate function curve (in the case of a function of two variables, the
volume between the function surfaces). The L2 norm of is similar but gives
more weight to large errors. The L norm of is simply the maximum value
of the error in the region.
As with other integral values, function norms are usually evaluated as a finite
sum. A discrete form of the Lp norm is given by
1/ p


p
Lp ( ) = (xi ) x (6.3)
i

Note that if the discrete L2 norm is used as a local measure of mesh element
error, merely subdividing an element will produce elements with a smaller
local error, even if the global error is unaffected, since the x are related to
the element size. The root mean square (RMS) error is a useful alternative
in this case. The RMS error is related to the L2 norm and is given by

i ( f (xi ) f (xi ))2 Wi

RMS = (6.4)
i Wi
where the Wi are weights assigned to each sample point. Typically each
weight is the area of the domain represented by the sample point, in which
case the RMS error is an area average of the L2 norm.

For a more detailed discussion of function norms, and of functional approx-

imation in general, see, for example, [180].

Radiosity and Realistic Image Synthesis 133

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.1 ERROR METRICS

Figure 6.1: Error measures: true versus approximate error.

Figure 6.1(a) shows approximations of a function f(x) using piecewise con-

stant and piecewise linear basis functions. The shaded areas between f(x) and
the approximation curves represent the approximation error. Figure 6.1(b) shows
two plots depicting the estimation of this error by the difference between the
original approximation and a higher-order approximation constructed using the
same nodal values. In the upper plot the error in a constant element approxi-
mation is estimated by comparing it to linear interpolation. In the lower plot,
linear and quadratic interpolation are compared.
This estimate of the error is typically evaluated over a local region of the
domain, for example, over an element when deciding whether it should be re-
placed by several smaller elements. A single value characterizing the magnitude
of the error function can be obtained using a function norm, such as one of the
Lp norms (see the box on page 133 for a discussion of function norms).

6.1.3 Residual of the Approximate Solution

The residual of the approximate solution provides another approach to charac-
terizing (x). The residual r(x) is obtained by substituting B (x) for B(x) in

Radiosity and Realistic Image Synthesis 134

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.1 ERROR METRICS

the original radiosity integral equation:

r(x) = B (x) E(x) S (x) G(x, x(x)) B (x) dA (6.9)

Clearly, if the residual is zero everywhere, then B (x) is an exact solution

to the integral equation and the actual error is also zero. Otherwise, the resid-
ual function is not identical to the actual error function. Strictly speaking, the
residual provides a global rather than local error estimate; a high residual in
a certain local region does not necessarily mean that B (x) is a poor approxi-
mation over that particular region. The residual in that region may be due to
inaccuracies elsewhere in the domain. However, in practice the residual is often
a useful indicator of local error introduced by the approximation. The residual
is expensive to compute because of the integration and is normally evaluated
only at selected points.

6.1.4 Error Based on the Behavior of the Kernel

As discussed in Chapter 3, the kernel B (x) G(x, x) of the radiosity integral
equation is also approximated by projecting it onto the selected basis functions.
Thus the error can also be characterized by examining the behavior of the ker-
nel function itself rather than the radiosity approximation. For example, while
computing the form factors one might recognize large local variations in the
geometric kernel, G(x, x), across certain pairs of elements. These gradients
might result from surfaces that are in close proximity or from shadow edges or
penumbra.
The geometric term, G(x, x), is a function of two points, typically on two
different elements. When it is determined that the basis functions cannot capture
the local variation of the geometric kernel itself, a decision can be made to
subdivide one or both of the elements in order to reduce the error introduced by
the mesh. This topic will be explored in much greater detail in Chapter 7 in the
context of hierarchical subdivision methods.

6.1.5 Image Based Error Metrics

The error metrics discussed in the previous sections are based on quantities
computed in the same domain as the radiosities, that is, the surfaces. However,
the final product of an image synthesis algorithm is actually an image comprised
of a finite number of pixels. If an image for a known viewpoint is the goal
of the global illumination solution, then an error metric that incorporates view
information can provide a valuable means of limiting and focusing computational
effort.

Radiosity and Realistic Image Synthesis 135

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.1 ERROR METRICS

Given a particular set of viewing parameters, the accuracy of the approxi-

mation for the visible surfaces is clearly the most critical, or the most important.
The importance of other surfaces depends on their contribution to the illumina-
tion of the visible surfaces. Smits et al. [220] incorporate a quantity measuring
importance in this case into an image-based error metric. Much like radiosity,
importance depends on surface geometry and reflectivity. It is propagated back
from the eye point into the environment, much as radiosity is propagated for-
ward from the light sources. Smits formulation and algorithm are described in
detail in Chapter 7.

6.1.6 Perceptually Based Error Metrics

The development of quantitative error metrics for radiosity [116, 120] is a crucial
step toward reliable algorithms. However, for image synthesis applications, it
is important to keep in mind that computing B(x) is not the ultimate goal. For
image synthesis, computing the radiosity, or radiance, is just one step in the
process of generating a visual experience.
As outlined in Chapter 1 and discussed in detail in Chapter 9, the dis-
play and perception of an image involve converting the radiance or radiosity to
screen values and finally to brightness as perceived by the eye and brain. These
transformations are nonlinear and highly complex. As a result, certain errors
may degrade perceived image quality to a greater or lesser degree than an error
measure based on photometric or radiometric quantities alone would indicate.
A perceptually based error metric would perhaps be used on the subjective
quantity of perceived brightness rather than radiosity. Low-order derivatives
might be incorporated into the metric to account for the sensitivity of the eye
to contrast. In the absence of such a metric, current attempts to incorporate
perception into the image synthesis process are largely heuristic or ad hoc. 1
Perceptually based error metrics for image synthesis remain an important re-
search topic.

6.2 Mesh Characteristics and Accuracy

An important goal when creating a mesh is to produce an approximation of
the required accuracy as inexpensively as possible. This requires understanding
how various characteristics of the mesh affect the accuracy of the approxima-
tion. These characteristics can be classified into four broad categories: mesh
density, element order, element shape and discontinuity representation. Before
discussing each of these in more detail, a concrete example will be presented

1
For exceptions and valuable discussions of this issue, see [238] and [250]. Subjective
brightness and image display are discussed at length in Chapter 9.

Radiosity and Realistic Image Synthesis 136

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.2: A radiosity image computed using a uniform mesh.

(see Figures 6.2 through 6.4) to illustrate the related issues of mesh quality,
approximation error and visual quality.

6.2.1 An Example
The images in this series compare a radiosity solution to a correct,, reference
image. The model is a simple L-shaped room illuminated by a single, diffusely
emitting area light.
The image in Figure 6.2 shows a radiosity solution on a uniform mesh of
linear elements performed using point collocation and rendered with Gouraud
interpolation. The solution was performed for direct illumination only (i.e.,
interreflected light was ignored). The form factors to the light source were
computed at each mesh node by shooting 256 rays to randomly selected points
on the light polygon.
The reference image (see Figure 6.3(a)) was computed by similarly eval-
uating the form factor to the light at the surface point visible at the center of
each image pixel. Since the contribution of the light source has been evaluated
similarly in both the radiosity and reference images, the difference between the
two images is essentially due to the approximation created by the element mesh.
Numerous artifacts introduced by the approximation are evident, including
blocky shadows (letter A of Figure 6.3(b)), missing features (letter B), Mach

Radiosity and Realistic Image Synthesis 137

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

(a) Reference image.

(b) Artifacts introduced by the approximation.

Figure 6.3:

Radiosity and Realistic Image Synthesis 138

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.4: Error image.

bands (letter C), inappropriate shading discontinuities (letter D), and unresolved
discontinuities (letter E).
Figure 6.4 is a visualization of the RMS error (defined in the box on function
norms on page 133) for each element. The error was computed by evaluating
the radiosity at 16 interior points across each element and comparing this to
the approximate value. The comparison shows absolute as opposed to relative
error. As apparent in this image, the error in the approximation is very unevenly
distributed over the mesh. Computational effort has been applied fairly equally
over the domain of the approximation, with very unequal contributions to the
resulting accuracy.
The example shown in these images will be referred to throughout this
chapter, to illustrate the discussion of meshing strategies. That discussion will
begin with the description of basic mesh characteristics in the following sections.

6.2.2 Mesh Density

Mesh density is determined by the size of the elements into which the domain is
subdivided. The density of the element mesh determines the number of degrees
of freedom available to the approximation. More elements provide more degrees
of freedom, which allows the approximation to follow the actual function more
closely and increases the degree to which small features like shadows can be
resolved.

Radiosity and Realistic Image Synthesis 139

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Linear elements with uniform subdivision.

Linear elements with higher density subdivision.

Linear elements with non-uniform subdivision.

Figure 6.5: Comparison of element subdivision strategies.

Radiosity and Realistic Image Synthesis 140

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.6: A radiosity solution with the mesh density increased four times over
the mesh used in Figure 6.2.

To illustrate, a one-dimensional function is approximated in Figure 6.5 using

a variety of subdivisions. In the coarsest subdivision (the topmost plot in Fig-
ure 6.5) the approximation follows the actual radiosity function closely where it
is almost linear, but diverges where the function changes slope quickly. Smaller
features of the function are missing entirely when they fall between nodes.
Evaluating the function at smaller intervals increases the accuracy of the
approximation, as shown in the second plot of Figure 6.5. However, the errors
remain unevenly distributed, meaning that the extra effort has been applied
inefficiently. Ideally, the domain should be subdivided more finely only where
it will improve the accuracy significantly, as in the third plot of Figure 6.5.
Similar observations apply to approximating the radiosity function. In the
example image (see Figure 6.2), inadequate mesh density results in elements that
are too large to capture shading detail accurately. This is particularly evident in
the staircase shadows cast by the table (letter A in Figure 6.3(b)), where the
size of the shading feature (the penumbra) is much smaller than the separation
between nodes.
Just as in the one-dimensional case, uniformly increasing the mesh density
improves the overall accuracy. The image in Figure 6.6 was produced using
a uniform mesh with four times as many elements as used in Figure 6.2. The

Radiosity and Realistic Image Synthesis 141

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.7: RMS error for the elements in figure 6.6.

quality is better and some of the artifacts are almost eliminated. However,
the corresponding error image in Figure 6.7 shows that the extra effort has been
applied inefficiently, with many elements subdivided where the error was already
negligible. This effort would have been better expended in further reducing the
error in the remaining problem areas, such as along the shadow boundaries.

6.2.3 Element Order and Continuity

Element order (i.e., the order of the basis functions defined over the element) also
has a direct effect on accuracy. Higher-order elements interpolate the function
using higher-order polynomials and use more information about the behavior
of the function, either by evaluating both values and derivatives at the nodes
or by evaluating the function at Additional nodes. Thus, higher-order elements
can follow the local variations in a function more closely than the same number
of lower-order elements. However, higher-order elements are generally more
expensive to evaluate. Thus, one approach is to use higher-order elements only
where the error is high and the extra effort is justified.
The type of basis functions used also affects the degree of continuity in the
approximation at element boundaries. Continuity of value and of lower-order
derivatives (C0, C1, C2, ...) is important because the human visual system is

Radiosity and Realistic Image Synthesis 142

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.8: Mach bands caused by first-derivative discontinuities at element

boundaries.

highly sensitive to relative spatial variation in luminance and its derivatives. 2

For example, the bright streaks along the wall in the example image (see
the closeup in Figure 6.8) correspond to discontinuities in the first derivative of
the radiosity approximation, which occur at the boundaries of linear elements.
The eye accentuates first derivative discontinuities, resulting in the perceptual
phenomenon known as Mach bands [189]. Although Mach bands can occur
naturally, they are distracting when they are incorrectly introduced by the ap-
proximation.
Since interpolation within elements is a linear sum of polynomials, the ap-
proximation is smooth (C) on element interiors. However, continuity at ele-
ment boundaries is not guaranteed. For Lagrange elements interpolation inside
or on the boundary of the element depends only on the function values at the
nodes belonging to that element. Because interpolation along boundaries uses
only nodes on the boundary, linear and higher order Lagrange elements guaran-
tee C0 continuity at element boundaries. (Constant elements are discontinuous

2
The notation Ck indicates that a function is continuous in all derivatives up to and
including k. C0 thus indicates that a function is continuous in value (i.e., there are no
sudden jumps in value), C1 that the function is continuous in slope (i.e., there are no
kinks), and C that the function is smooth (i.e., continuous in all derivatives).

Radiosity and Realistic Image Synthesis 143

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.9: A combination of rectangular and triangular elements used to fill a

complicated geometry.

in values at element boundaries). Hermite elements can provide a higher degree

of continuity with fewer nodes by interpolating nodal derivatives as well as val-
ues. C1 continuity, for example, requires interpolating the gradient at boundary
nodes.
For radiosity, providing continuity greater than C0 is motivated primarily by
perceptual issues. Thus, elements with this property, such as the Clough-Tocher
element used by Salesin et al. [203], have been applied mainly to the rendering
stage. These will be discussed in detail in Chapter 9.

6.2.4 Element Shape

Elements should provide well-behaved, easily evaluated basis functions. For this
reason, standard element shapes consist of simple geometries, typically triangles
and rectangles. These standard shapes can be mixed to subdivide a complicated
geometry more efficiently (see figure 6.9) or to provide a transition between
regions of differing mesh density (see Figure 6.10).
Isoparametric elements allow the standard shapes to be mapped to more
general geometries, using the basis functions to interpolate geometric location
as well as radiosity values (see section 3.8). The bilinear quadrilateral
element is a common example. Higher order isoparametric elements can be
mapped to curved geometries. The parametric mapping must be invertible,
which places some restrictions on the element shape. Concavities and extra
vertices are normally to be avoided.
If Gouraud shading is to be used for rewndering, anumber of special problems
relating to element shape must be avoided. For example, Gouraud interpolation

Radiosity and Realistic Image Synthesis 144

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.10: Use of triangular elements to provide transition between regions of

high and low mesh density.

over a concave element can generate discontinuities in value, as shown in Fig-

ure 6.11. The problems relating to element shape and Gouraud interpolation
will be discussed in Chapter 9.

Aspect Ratio
Element shape affects the efficiency with which the mesh samples the radiosity
function [18]. To make the most efficient use of a given number of nodes, the
nodes should be evenly distributed over the domain of the function (assuming
that the behavior of the function is unknown). A distribution of nodes that is
denser in one direction than in another is inefficient, since if elements are subdi-
vided far enough to make sampling adequate in the sparse direction, sampling in
the denser direction will be greater than necessary. Poor element shape can also
affect the accuracy or efficiency of numerical integration of the form factors.
A reasonably uniform distribution of nodes can be obtained by requiring
elements to have as high an aspect ratio as possible. The aspect ratio is defined
as the ratio of the radius of the largest circle that will fit completely inside the
element to the radius of the smallest circle that will fit completely outside the
element (see Figure 6.12) [18]. This ratio should be as close to 1 as possible.
However, if the behavior of the function is known, anisotropic sampling

Radiosity and Realistic Image Synthesis 145

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.11: A concave element, with a discontinuity introduced by Gouraud

shading (closeup of letter D in Figure 6.3 (b)).

may be more efficient than a uniform distribution, since the function changes
more slowly in the direction orthogonal to the gradient and thus fewer samples
are required along that direction. Elements with a lower aspect ratio may be
more efficient in this case, if oriented correctly.

In a review of surface triangulation techniques, Schumaker [207] provides an

excellent example of the effect of element orientation. A function surface is first
approximated by triangulating regularly spaced nodes. However, the approxi-
mation is more accurate when the nodes are connected to form triangles with
poor aspect ratios (see Figure 6.13). Schumaker discusses how approximation
accuracy can be incorporated into the quality metric used by the triangulation
algorithm. The approximation of a surface in 3-space is analogous to the ap-
proximation of the radiosity function and the surface approximation literature is
thus a fruitful source for meshing algorithms.

Form factor algorithms may make assumptions that are violated by elements
with poor aspect ratios. For example, in the adaptive ray casting algorithm
described by Wallace et al. [247], (see Chapter 4), elements or pieces of elements
(delta-areas) are approximated as disks. The accuracy of the disk approximation
decreases with decreasing element aspect ratio.

Radiosity and Realistic Image Synthesis 146

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.12: The effect of aspect ratio on sampling density. A low aspect ratio
tends to produce an anisotropic sampling density.

Mesh Grading
A nonuniform, or graded, mesh may be required to distribute the error evenly in
the approximation. In a well-graded mesh, element size and shape vary smoothly
in the transition between regions of higher and lower density (see Figure 6.14).
Abrupt changes in size or shape will often cause visible irregularities in shading,
since neighboring elements will approximate the function with slightly different
results. Because of the eyes sensitivity to contrast, such differences may be
visible even when the error for both elements is within tolerance.

Mesh Conformance
It is important that adjacent elements conform across shared boundaries, as
shown in Figure 6.15. Discontinuities in value created by nonconforming ele-
ments can cause distinctive image artifacts. Conforming elements share nodes

Radiosity and Realistic Image Synthesis 147

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.13: (a) A surface approximation based on a regular subdivision. (b) The
same surface approximated using elements with poor aspect ratios but oriented
so as to provide a better approximation, (after Schumaker, 1993).

Figure 6.14: The top mesh shows a poorly graded transition between regions of
high and low mesh density. The bottom well-graded mesh provides a smoother
transition.

Radiosity and Realistic Image Synthesis 148

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.15: Conforming versus nonconforming elements.

along boundaries, thus ensuring C0 continuity at the boundaries. In practice, this

means that T-vertices (defined in Figure 6.15) must be avoided. Conformance
is not an issue for constant elements, which are often used during the radiosity
solution, but it is critical for rendering, where linear or higher-order elements
are normally used.

6.2.5 Discontinuities
If light sources have a constant emission value (or are C ) across the surface
and if changes in visibility are ignored, then the radiosity function B(x) across a
receiving surface will be continuous in all derivatives (i.e., it will also be C ).
This is evident from the form factor kernel, G(x, x), which itself is C except
where the visibility term V(x, x) changes from one to zero, and at singularities
where x and x meet and the denominator goes to 0.
If changes in visibility (i.e., shadows) are included, the radiosity function can
contain discontinuities of any order [121]. Discontinuities in value and in the first
and second derivatives are the most important, since these often provide visual
cues to three-dimensional shape, proximity, and other geometric relationships.
Much of the image processing performed by the eye involves enhancing such
discontinuities and, as a result, the failure to reproduce discontinuities correctly
can degrade image quality dramatically.

Radiosity and Realistic Image Synthesis 149

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.16: Shadows can cause 0th, 1st, 2nd, and higher-order discontinuities
in the radiosity function across a surface.

Value Discontinuities
Discontinuities in the value of the radiosity function occur where one surface
touches another, as at letter a of Figure 6.16. In Figure 6.17 the actual radiosity
function is discontinuous in value where the wall passes below the table top. The
shading should thus change abruptly from light to dark at the boundary defined
by the intersection of the wall and the table. Unfortunately, the boundary falls
across the interior of elements on the wall. Instead of resolving the discontinuity,
interpolation creates a smooth transition and the shadow on the wall appears to
leak upwards from below the table top.
Incorrectly resolved value discontinuities can also cause light leaks, in
which interpolation across one or more discontinuities causes light to appear
where shadow is expected. In Figure 6.18, for example, elements on the floor
pass beneath the wall dividing the room. The resulting light leak gives the
incorrect impression of a gap between the wall and the floor. The problem
is compounded when these elements incorrectly contribute illumination to the
room on the left, which is totally cut off from the room containing the light
source.

Derivative Discontinuities
Discontinuities in the first or second derivative occur at penumbra and umbra
boundaries (letter b of Figure 6.16), as well as within the penumbra. When mesh
elements span these discontinuities, interpolation often produces an inaccurate
and irregular shadow boundary. The staircase shadows in Figure 6.2 are an
example.

Radiosity and Realistic Image Synthesis 150

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Figure 6.17: Failure to resolve a discontinuity in value. This is a closeup of the

radiosity solution shown in Figure 6.2.

Figure 6.18: A light leak caused by failure to resolve discontinuities in value

where the dividing wall touches the floor. The dividing wall completely separates
the left side of the room from the right side, which contains the light source.

Radiosity and Realistic Image Synthesis 151

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.2 MESH CHARACTERISTICS AND ACCURACY

Singularities in the first derivative can also occur, as at letter c of Figure 6.16
where the penumbra collapses to a single point. Tracing along the line of
intersection between the two objects, an instantaneous transition from light to
dark is encountered at the corner point. The first derivative is infinite at that
point, although the function is continuous away from the boundary.
The correct resolution of discontinuities requires that they fall along ele-
ment boundaries, since the approximation is always C on element interiors.
Thus, discontinuity boundaries must either be determined before meshing or the
mesh must adapt dynamically to place element edges along the discontinuities.
Since discontinuities may be of various orders, interpolation schemes that can
enforce the appropriate degree of continuity at a particular element boundary are
also required. Techniques for finding and reconstructing discontinuities will be
discussed in detail in Chapter 8.

Continuity at Geometric Boundaries

Discontinuities in value occur naturally at boundaries where the surface nor-
mal is discontinuous, such as where the edge of the floor meets a wall. Such
discontinuities are normally resolved automatically, since surfaces are meshed
independently.
A problem can occur, however, if the boundaries of the primitives gener-
ated by a geometric modeler do not correspond to discontinuities in the surface
normal. For example, curved surfaces will often be represented by collections
of independent polygonal facets. If the facets are meshed independently, adja-
cent elements will often be nonconforming across facet boundaries, and shading
discontinuities will result, as shown in Figure 6.19. It is easiest to maintain
conformance in this case if the connectivity of the facets is determined prior
to meshing and the surface is meshed as a single unlit (see Figure 6.20). This
approach is used by Baum et al. [18] for the special case of coplanar facets. Bet-
ter yet, the radiosity implementation should allow the user to enter the faceted
surfaces as a topologically connected primitive such as a polyhedron.

6.3 Automatic Meshing Algorithms

With a better understanding of how various mesh attributes affect the accuracy
of the solution, it is now possible to discuss automatic meshing strategies. A
taxonomy of automatic meshing algorithms is shown in Figure 6.21. 3

3
Although user intervention can be helpful in constructing a mesh, the discussion
in this chapter will be limited to automatic mesh generation. Meshes for engineering
applications are still often constructed with some interactive help, but good results require
an experienced user who understands the underlying principles of the analysis. In image

Radiosity and Realistic Image Synthesis 152

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.19: The polygonal facets representing the curved surface in this image
were meshed independently. The resulting elements are nonconforming at the
facet boundaries, causing shading discontinuities.

Figure 6.20: The facets in this image were connected topologically prior to
meshing, and the surface was meshed as a unit.

Radiosity and Realistic Image Synthesis 153

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Automatic Meshing Strategies

No Knowledge of Function Knowledge of Function

Uniform Non-Uniform A Priori A Posteriori

r-refinement h-refinement p-refinement remeshing

Figure 6.21: A taxonomy of automatic meshing strategies.

Meshing algorithms can be broadly classified according to whether or not

they use information about the behavior of the function to be approximated. Al-
though obtaining an optimal mesh normally requires such knowledge, in practice
some degree of meshing without it is almost always necessary. Meshing in this
case generally means subdividing as uniformly as possible (although subdividing
complex geometries may require a nonuniform mesh). Algorithms for producing
a uniform mesh are described in Chapter 8.
Meshing techniques that use knowledge of the function can be characterized
as either a priori or a posteriori [211]. A priori methods specify all or part of
the mesh before the solution is performed. Discontinuity meshing, in which
discontinuity boundaries associated with occlusion are determined prior to the
solution based on purely geometric considerations, is an a priori method. A
priori algorithms, including discontinuity meshing, are discussed in Chapter 8.
A posteriori algorithms determine or refine the mesh after the solution has
been at least partially completed. An initial approximation is obtained using
a uniform or other mesh determined a priori. The mesh is then refined in
regions where the local error is high, using information provided by the initial
approximation of the function, such as the gradient, to guide decisions about
element size, shape, and orientation. A posteriori meshing strategies are the
subject of the remainder of this chapter.

6.3.1 A Posteriori Meshing

A posteriori meshing algorithms common to finite element analysis can be cat-
egorized as follows [211]:

synthesis the analysis of illumination is typically not the users primary task, and the
detailed specification of a mesh is intrusive and often beyond the users expertise.

Radiosity and Realistic Image Synthesis 154

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

r-refinement: reposition nodes

h-refinement: subdivide existing elements

p-refinement: increase polynomial order of existing elements

remeshing: replace existing mesh with new mesh

Each of these approaches addresses one or more of the basic mesh characteristics
discussed earlier: mesh density, basis function order, and element shape. Radios-
ity algorithms have so far relied almost exclusively on h-refinement. However,
the other approaches will also be briefly described here, partly to indicate pos-
sible directions for radiosity research. See Figure 6.22 for illustrations of these
approaches.

R-refinement
In r-refinement the nodes of the initial mesh are moved or relocated during
multiple passes of mesh relaxation. At each pass, each node of the mesh is
moved in a direction that tends to equalize the error of the elements that share the
node. (See section 8.4 for a basic algorithmic approach to moving the vertices.)
The function is then reevaluated at the new node locations. Relaxation can
continue until the error is evenly distributed among the elements.
R-refinement has the advantage that the mesh topology is not altered by
the refinement, which may simplify algorithms. It generates an efficient mesh,
in that it minimizes the approximation error for a given number of elements
and nodes. On the other hand, r-refinement cannot guarantee that a given error
tolerance will be achieved. Since the number of elements is fixed, once the error
is evenly distributed it cant be lowered further. Also, care must be taken during
relocation not to move nodes across element boundaries. It may furthermore be
difficult to maintain good element shape near fixed boundaries.

H-refinement
In h-refinement, the local error is decreased by increasing the density of the
mesh; elements with a high error are subdivided into smaller elements. (The
h refers to the symbol commonly used to characterize element size in finite
element analysis.) The function is then evaluated at the new nodes. Since
new nodes are added, the approximation error can be made as small as desired
(although this may not always be practical). Elements and nodes can also be
removed in regions where the approximation error is lower than necessary. In
some h-refinement algorithms, refinement does not require reevaluation of the
function at existing nodes.

Radiosity and Realistic Image Synthesis 155

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.22: Basic a posteriori meshing strategies.

However, the inability to move existing nodes restricts the ability of h-

refinement to reduce error by adjusting element shape or orientation. As a
result, h-refinement can be inefficient, in that more elements than necessary
may be needed to reach a desired accuracy. Special handling is required to
maintain continuity between elements subdivided to different levels of refine-
ment. H-refinement algorithms must also pay careful attention to mesh grading.
Radiosity implementations have relied almost exclusively on a variety of h-
refinement algorithms. These are expanded on in section 6.3.2. Babuska et al.
[15] provide a valuable source for h-refinement and p-refinement approaches
used in engineering applications.

Radiosity and Realistic Image Synthesis 156

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

P-refinement
In p-refinement, the approximation error is reduced by increasing the order of the
basis functions for certain elements. (The symbol p refers to the polynomial
order of the basis functions.) New nodes are added to the affected elements,
but the element shape and the mesh topology are not otherwise changed. In
contrast to h-refinement, the number of elements in the mesh is not increased,
which limits computational costs in some ways. However, the function must be
evaluated at additional nodes, and the higher-order basis functions can be more
expensive to compute. As with h-refinement, the ability to change element
shape or orientation is restricted. Care must also be taken to maintain continuity
between adjacent elements having basis functions of different orders.

Remeshing
Remeshing algorithms modify both the node locations and mesh topology; in
essence the existing mesh is completely replaced. This allows complete flexibil-
ity of element shape and orientation, as well as the ability to decrease arbitrarily
the approximation error. Following remeshing, the function must be reevaluated
at the nodes of the new mesh.

Hybrid Methods
Hybrid refinement methods can be constructed by combining the above basic
approaches. For example, r-refinement works best when it begins with a mesh
that is reasonably well refined, since relaxation cannot reduce the local error
beyond the minimum achievable with the initial number of nodes. A potentially
useful hybrid strategy might thus use h-refinement to achieve a reasonable initial
sampling density and r-refinement to more evenly distribute the approximation
error.

6.3.2 Adaptive Subdivision: H-refinement for Radiosity

Almost all a posteriori meshing algorithms for radiosity have used h-refinement.
This approach, commonly called adaptive subdivision in radiosity applications,
follows the basic outline of an a posteriori method: a solution is computed on a
uniform initial mesh, and the mesh is then refined by subdividing elements that
exceed some error tolerance.
Figure 6.23 illustrates the improvement provided by adaptive subdivision
over the uniform mesh approximation shown at the beginning of the chapter in
Figure 6.2. The image quality is improved compared to that provided by the
high uniform mesh resolution of Figure 6.6, while using the same number of

Radiosity and Realistic Image Synthesis 157

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.23: Adaptive subdivision. Compare to Figure 6.2.

Figure 6.24: Error image for adaptive subdivision. Compare to Figures 6.4 and 6. 7.

Radiosity and Realistic Image Synthesis 158

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Adaptive_Subdivision( error_tolerance ) {
Create initial mesh of constant elements ;
Compute form factors ;
Solve linear system ;
do until ( all elements within error tolerance
or minimum element size reached ) {
Evaluate accuracy by comparing adjacent element radiosities ;
Subdivide elements that exceed user-supplied error tolerance ;
for ( each new element ) {
Compute form factors from new element to all other elements ;
Compute radiosity of new element based on old radiosity values ;
}
}
}

Figure 6.25: Adaptive subdivision pseudocode.

elements. The corresponding error image is shown in Figure 6.24. Note that
even with adaptive subdivision, the error remains high for elements lying along
the shadow boundary.
Cohen et al. [61] first applied adaptive subdivision to radiosity meshing,
using the algorithm outlined in the pseudocode in Figure 6.25. For clarity, this
outline ignores the hierarchical nature of the adaptive subdivision algorithm,
which will be discussed in detail in the following chapter. For now, note only
that new nodes created by adaptive subdivision have their radiosities computed
using the approximation B (x) obtained during the initial solution. As a result,
it is not necessary to recompute radiosities for existing nodes when an element
is subdivided.
Many variations of this basic approach have been developed differing pri-
marily in how they estimate error, and how elements are subdivided. The fol-
lowing sections will survey adaptive subdivision algorithms and how they have
addressed these two issues.

6.3.3 Error Estimation for Adaptive Subdivision

Heuristic and Low-Order Subdivision Criteria

Many algorithms subdivide according to a discrete approximation to one or
more of the error norms described in section 6.1. For example, Cohen et al.

Radiosity and Realistic Image Synthesis 159

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.26: Cohens subdivision criterion based on the difference between nodal
values results in the subdivision of element A, although linear interpolation pro-
vides a good approximation in this case. (Added nodes due to the subdivision
are indicated by hollow circles.) The local minimum at D is also missed.

[61] compare the radiosities of an element and its neighbors . If these differ in
value by more than a user-specified tolerance, the element is subdivided. For the
constant elements, this is essentially equivalent to estimating the local error by
comparing the piecewise constant approximation to linear interpolation through
the same nodes and nodal values.
For rendering, however, Cohen uses linear interpolation. With respect to
linear interpolation, this subdivision criterion is better characterized as a heuristic
designed to produce smaller elements in regions where the radiosity is highly
variable. This heuristic usually produces acceptable results, although it tends to
oversubdivide where the gradient is high but constant. Since linear interpolation
is a reasonable approximation for this case, subdividing has little effect on the
error (see Figure 6.26.)
This heuristic may also fail to identify elements that should be subdivided.
In Figure 6.26 the nodes bounding an element containing a local minimum (letter
D in the figure) happen to have almost the same value. The heuristic fails in this
case, since the nodal values alone do not provide enough information about the
behavior of the function on the element interior. This difficulty is common to all
types of error estimators and algorithm make efforts of varying sophistication
to characterize the function between nodes.
For example, Vedel and Puech [242] use the gradient at the nodes as well as
the function value. Elements are subdivided if the gradients at the element nodes
vary by more than a certain tolerance (see Figure 6.27). This criterion avoids
subdividing elements unnecessarily where the gradient is high but constant (letter

Radiosity and Realistic Image Synthesis 160

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.27: Gradient-based subdivision criterion. Nodes added due to adaptive

subdivision are indicated by hollow circles.

A of Figure 6.27). It may also detect a local minimum within an element whose
nodal values happen to be similar (letter B of Figure 6.27). However, the
criterion is not foolproof. Letter C of Figure 6.27 shows a situation that might
occur when a penumbra falls entirely within an element. The gradients at the
nodes spanning the element happen to be equal and the element is not subdivided.
A more stringent criterion can be constructed that uses both the nodal values
and gradients. An algorithm based on this approach might first compare gradi-
ents at the nodes. If the gradients vary by too much, the element is subdivided.
Otherwise, the gradient at the nodes is compared to the slope of the plane deter-
mined by the nodal values. If they are inconsistent, the element is subdivided.
This criterion correctly identifies the element at letter C in Figure 6.27, although
it does not identify the case indicated by the letter D.

Higher-Order Subdivision Criteria

Vedel and Puech describe a test-bed system that estimates local error based on
a higher-order (bicubic) interpolant over rectangular elements [242]. The local
error estimate is provided by the L2 norm of the difference between bilinear
and bicubic interpolation over an element (see Figure 6.28). This integral is
evaluated in closed form as a function of the radiosities and gradients at the
element nodes.
In comparing the higher-order error estimate to valueand gradient-based
criteria for several simple cases, Vedel and Puech find the bicubic interpola-
tion estimate to be the most accurate of the three when the radiosity is slowly
varying. However, it fails to identify elements across which the radiosity is

Radiosity and Realistic Image Synthesis 161

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.28: Estimation of error by comparing linear and cubic interpolation.

The gray area represents the estimated error. Nodes that will be added due to
adaptive subdivision are indicated by hollow circles.

changing very rapidly, for example, where the element contains a sharp shadow
boundary. In this case, the simple comparison of values does better, although
as expected, it tends to oversubdivide in other regions. Vedel and Puech con-
clude that the complexity of the bicubic method is not justified, and they suggest
simply comparing both radiosities and gradients.

Estimation Using the Residual

All of the methods described so far ultimately fail at some point because the
nodes can never be relied on to completely characterize the behavior of the
function elsewhere in the element. A local minimum or maximum can fall
entirely within an element without affecting the function or its derivatives at the
nodes, as is the case for the local minimum at letter D in Figure 6.26. Small
shadows are often missed for this reason, with one common result being the
appearance of floating furniture in radiosity images (see the artifact labeled B in
Figure 6.3(b).)
In such cases, the only solution is to evaluate the function inside the element.
One approach is to evaluate the radiosity equation at one or more points within
the element and compare the results to the interpolated values. This is equivalent
to estimating the residual, (described in section 6.1.3.)
Lischinski et al. [153] have used this technique for one-dimensional elements
in a flatland radiosity implementation. In [154] they generalize this approach
to two-dimensional elements by evaluating the radiosity at the centroid of the
element and comparing it to the interpolated value at the same location (see

Radiosity and Realistic Image Synthesis 162

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.29: Estimation of error by computing the residual at element centers.

The hollow circles represent interpolated values and the solid circles the com-
puted value. The residual is the difference between the two.

Figure 6.29). The error at the centroid is assumed to be the maximum error
for the element. This approach can thus be viewed as estimating the L norm.
This technique is independent of the interpolation order (quadratic interpolation
was used by Lischinski et al.).
Of course, evaluating the error at the centroid is not guaranteed to catch
every case. In Lischinskis implementation, mesh boundaries corresponding to
discontinuities in the radiosity function or its derivatives are specified a priori.
Thus, a posteriori adaptive subdivision is required to refine the mesh only within
regions over which the radiosity function is relatively smooth and well behaved,
in which case checking the error at the centroid will generally produce good
results.
Campbell [42] describes a more systematic approach to determining the
behavior of the function on the interior. This is particularly useful when the
radiosity function cannot be assumed to be smooth. Campbells criterion for
subdivision uses the difference between the maximum and minimum radiosities
over the entire element, not just at the nodes. Elements that require subdivision
are split perpendicularly to the line connecting the maximum and minimum
points. Thus, Campbell's algorithm depends on a systematic search for the
extrema, which is achieved using standard optimization techniques.
Since Campbell's algorithm computes shadow boundaries a priori, it can
identify fully lit regions and treat them differently from penumbra regions, which
are more complex. For fully lit regions Campbell computes the gradient at the
nodes analytically by differentiating the point-polygon form factor equation.
This allows the use of optimization techniques that take advantage of gradient

Radiosity and Realistic Image Synthesis 163

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING

information to accelerate the search. In addition, for a fully lit region there
can be only one local maximum on the interior of the element due to a given
constant diffuse source.
For regions within the penumbra the gradient cannot be computed analyti-
cally and no assumptions can be made about the number of local extrema. In
this case global optimization is performed using the Multistart method. A grid
is laid over the region and the function is evaluated at a random point inside
each grid cell. Cells whose neighbors are either all greater or all lesser in value
than the cell itself provide the starting point for local optimization.
None of the error criteria that have been described in these sections can guar-
antee that small features will be found. This is one advantage of discontinuity
meshing (discussed in Chapter 8), which locates critical shading boundaries a
priori based on the model geometry. Within regions bounded by discontinuities,
the radiosity function is reasonably well behaved, and simple error estimators
are more reliable.

Computing the Gradient

A number of error estimators require the gradient of the radiosity at the nodes.
Campbell points out that the analytic expression for the form factor between
a differential area and a constant, unoccluded polygonal element is continuous
and differentiable [42]. The expression can thus be symbolically differentiated
to provide an analytic formula for the gradient at unoccluded nodes. However,
the gradient is difficult or impossible to compute analytically in the presence of
occlusion and is actually undefined at certain discontinuity boundaries.
Numerical differencing can also be used to compute partial derivatives. If the
nodes fall on a regular grid, a first-order estimate of the partial derivative along
grid lines can be made by comparing a nodal value with that of its neighbors.
This estimate can be computed using forward or backward differencing, given
by
B B(x i ) B(x i 1 )
= (6.10)
x x i x i 1

where xi and xi1 are neighboring nodes. Central differencing can also be used,
given by
B B(x i +1 ) B(x i 1 )
= (6.11)
x x i +1 x i 1

If the mesh is irregular, the tangent plane at the node can be estimated using a
least-squares fit of a plane to the values at the node and its immediate neighbors.
(The contouring literature is a good source for techniques of this kind [257].)
The accuracy of these techniques depends on the spacing between nodes.

Radiosity and Realistic Image Synthesis 164

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

Figure 6.30: Two different subdivisions of the same element. The upper right
subdivision does not reduce the overall error of the approximation.

Another option is to perform extra function evaluations at points in the

neighborhood of the node. For example, Salesin et al. [203] use quadratic
triangular elements having nodes at the midpoint of each edge. To compute the
gradient, a quadratic curve is fit to the values of the three nodes along each edge.
The tangents of the two parabolas intersecting a given corner node determine a
tangent plane at the node, and thus the gradient. This technique is described in
more detail in section 9.2.2.
Ward describes a technique useful for methods that compute the irradiance
at a point by sampling the visible surfaces over the entire hemisphere above
the point, as in the hemicube [253]. The method was developed in the context
of Wards Radiance lighting simulation system, which does not use radiosity.
However, a full-matrix (gathering) radiosity solution using the hemicube or sim-
ilar method for computing form factors is equally amenable to Ward's technique,
although as yet no radiosity implementations have taken advantage of it.

6.3.4 Deciding How to Subdivide

Identifying elements that require subdivision is only the first step. The goal
in identifying elements for subdivision is to reduce the local error for those

Radiosity and Realistic Image Synthesis 165

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 6. RENDERING
6.3 AUTOMATIC MESHING ALGORITHMS

elements, but the actual reduction in error will depend~ on how the elements are
subdivided. For example, in Figure 6.30, two ways of subdividing an element
are compared. In one case, the error is reduced significantly, while in the other
it is not reduced at all.
Subdividing intelligently requires some estimate of the behavior of the func-
tion inside the element. Campbells optimization approach, described in the
previous section, is one of the few algorithms that attempts to obtain and use
this information. Campbell searches for the maximum and minimum points
on the interior of the element. The element is then subdivided on a bound-
ary perpendicular to the line connecting the maximum and minimum points.
The flexibility required to subdivide in this way places demands on the actual
subdivision algorithm. Campbell chooses a BSP-tree based approach for this
reason.
Airey [5] and Sturzlinger [227] note a useful technique for the special case
of subdividing rectangles into triangles. The edge created to split the rectangle
should connect the nodes with the most similar radiosities, since this produces
the greatest reduction in the variation between the nodes of each of the resulting
elements. Schumaker [207] discusses a generalization of this approach in which
the behavior of the approximation is incorporated into the quality metric used
during triangulation of the set of nodes. This more general approach has not yet
been applied to radiosity, however.
In the absence of knowledge about the function behavior, the best that can
be done is to subdivide uniformly. This is the approach taken by most existing
adaptive subdivision algorithms. The resulting subdivision will depend on the
particular subdivision algorithm. One common approach is to subdivide ele-
ments into four similarly shaped elements, generating a quadtree subdivision
hierarchy. If the elements are triangles, another approach is to subdivide ele-
ments, by inserting nodes and adding new edges. These and a wide variety of
other subdivision algorithms are surveyed in Chapter 8, with some discussion
of how they can be applied to adaptive subdivision.

Radiosity and Realistic Image Synthesis 166

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS

Chapter 7

Hierarchical Methods
The meshing strategies surveyed in the previous chapter are designed to reduce
the computational cost of the radiosity solution by minimizing the number of
elements in the mesh. The solution cost depends strongly on the number of
elements, since solving the discrete radiosity equation requires computing an
interaction between every pair of elements. Thus, the cost of the radiosity
solution appears to be inherently O(n2) in the number of elements. Each of these
O(n2) relationships involves evaluating the form factor integral (the subject of
Chapter 4) and is thus expensive to compute. Hence, the goal of the meshing
strategies outlined in the previous chapter is to minimize the number of elements
while maintaining accuracy.
The subject of this chapter is an alternative approach to reducing the com-
putational cost of the radiosity algorithm. This approach keeps the same number
of elements, instead attempting to reduce the number of individual relationships,
or form factors, that have to be computed. For example, two groups of elements
separated widely in space might reasonably have the total interaction between all
pairs of individual elements represented by a single number computed once for
the entire group. Attaining this goal involves developing a hierarchical subdi-
vision of the surfaces and an associated hierarchy of interactions. The hierarchy
will provide a framework for deriving interactions between groups of elements,
which will result in computing many fewer than O(n2) interactions. In fact,
it will turn out that only O(n) form factors are required to represent the linear
operator K to within a desired error tolerance.
This chapter is divided into three major sections. The first two sections
describe hierarchical subdivision techniques that minimize the number of form
factors to be computed by grouping elements together. The first section assumes
that constant basis functions have been selected to approximate the radiosity
function. The section begins with a description of a two-level, patch-element
hierarchy and continues with the generalization of the basic hierarchical ap-
proach, resulting in an O(n) algorithm. The second section then describes an
alternate way of approaching the same goal, in which the hierarchical algorithms
are derived in terms of hierarchical basis functions. One such class of basis func-

Radiosity and Realistic Image Synthesis 167

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.1 A PHYSICAL EXAMPLE

Figure 7.1: Room with a desk and wall.

tions, known as wavelets, will be used to represent the kernel of the radiosity
function at variable levels of detail. This formulation will provide a framework
for incorporating higher order basis functions into hierarchical algorithms.
The third section of the chapter introduces an adjoint to the radiosity equa-
tion that allows the importance of a particular element to the final image to
be determined. The combined radiosity equation and its adjoint will provide
yet another means to reduce greatly the computational complexity for complex
environments, in the case where only one, or a few, views are required.

I. Hierarchical Subdivision

7.1 A Physical Example

The basic physical intuition behind hierarchical solution methods is straightfor-
ward. Imagine a room containing a table on which are placed several small
objects (see Figure 7.1). Light reflected from the table top contributes some
illumination to the wall. Intuitively, however, the shading of the wall does not
depend significantly on the small details of the illumination leaving the table
top. If the objects on the table are rearranged so that the shadows on the table
are slightly altered, the shading of the wall does not change significantly.
Representing the contribution of the table top to the wall by a single aver-
age value will give a similar result to computing individual contributions from
many small elements on the table top. This is because ideal diffuse reflection
effectively averages the light arriving over an incoming solid angle. If the solid
angle is not too great, as is the case when a source is far away relative to its

Radiosity and Realistic Image Synthesis 168

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.1 A PHYSICAL EXAMPLE

size from a receiver, it is reasonable to average the source radiosity before the
integration step.1 For radiosity, this means replacing several form factors with
a single form factor.
On the other hand, when the table top is eventually rendered in an image,
its shading must capture all the details of the light and shadows falling on it.
These details will be lost if the effect of illumination arriving at the table is
averaged over a large area rather than computed in detail for smaller regions.
Thus, the radiosity of a particular surface or group of surfaces will have to be
represented at at least two levels of detail: coarsely when the surface acts as a
source of illumination and more finely when it acts as a receiver. There is no
inherent limitation to two levels, and these notions will be extended in a later
section to a generalized hierarchical representation.

7.2 Two-Level Hierarchy

The first hierarchical algorithm for radiosity was developed by Cohen et al. [61].
It provides two levels of hierarchy and is based directly on the above-noted
distinction between sources and receivers. Surfaces are subdivided coarsely into
patches to represent the surface when it acts as a source of illumination to other
surfaces. The same surfaces, when receiving illumination, are represented by
a finer subdivision of the patches into elements. Images are rendered using
the element mesh. The element mesh can be subdivided adaptively as more
information about the illumination of the surface becomes known to achieve an
appropriate element size, but the patch subdivision is specified a priori by the
user. The algorithm assumes constant basis functions.
The steps in the two-level hierarchical algorithm are as follows (correspond-
ing to the numbers in Figure 7.2):

1. Divide the surfaces into m patches and n smaller elements (m << n),
where each patch is composed exactly of the union of some subset of
the elements. The patches will act as the sources or shooters and the
elements will act as the receivers or gatherers. The patches are indexed
by i or j and the elements by q.
2. Compute the m n form factors from each element to each patch. A
single entry from element q to patch j would be Fqj.
3. Compute the m m form factors between pairs of patches directly from
the element-to-patch form factors by summing the form factors for the

1
A similar observation enables the use of a local reflection model to approximate the
complex interactions of light with the microscopic geometry of a surface.

Radiosity and Realistic Image Synthesis 169

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.2 TWO-LEVEL HIERARCHY

Figure 7.2: Two-level hierarchical algorithm.

elements q belonging to patch i, weighted by the areas of the elements:

F
Aq
Fij = qj (7.1)
qi
Ai
4. Solve the m m system of equations using Gauss-Seidel to get the patch
radiosities:
m
Bi = Ei + i B F
j=1
j ij (7.2)

5. Back solve for the element radiosities. This is accomplished by plugging

the element-to-patch form factors computed in step 2 and the patch radiosi-
ties, Bj, from step 4, into the radiosity equation expressing the element
radiosity as the sum of contributions from each patch:
m
Bq = Eq + q B F
j=1
j qj (7.3)

Radiosity and Realistic Image Synthesis 170

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.3 THE K MATRIX

Figure 7.3: Quadtree surface subdivision.

6. If the difference in radiosity between neighboring elements is too high,

elements can be further subdivided. New patch-to-element form factors are
computed only for the new elements, and the new element radiosities can
be computed directly from the original patch radiosities without solving a
new system (i.e., the original m patches are intact). This last step can be
repeated adaptively as needed.

In practice, recursive adaptive subdivision is achieved by constructing a

quadtree2 of elements. If the original surfaces are initially divided into quadri-
laterals and triangles (see Figure 7.3), there is a simple operation that splits
each simple shape along each of its parametric midpoints, resulting in four new
smaller elements. The mechanics are discussed in greater detail in Chapter 8.
Following the solution, the radiosity function is approximated by the n
elements. However, the solution has required computing only m n form
factors, where m is much less than n.

7.3 The K Matrix

In Chapter 3 the discrete radiosity system was derived, resulting in the system
of linear equations KB = E, where the operator K is a matrix that represents
the interrelationships between nodes in the finite element approximation. It will

2
A quadtree is the two-dimensional analog to a binary tree (i.e., each node in the tree
has four children and one parent).

Radiosity and Realistic Image Synthesis 171

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.3 THE K MATRIX

be useful in what follows to understand the relationship between the physical

intuition developed in section 7.1 and structure in the K matrix.
In the two-level hierarchy of the previous section, grouping elements to-
gether into patches allows every element to compute only a single form factor
to each patch (rather than to each element), effectively reducing the number of
entries to be computed in the matrix K. The two-level hierarchy assumes that,
in effect, a number of form factors on the same row of K are closely related.
In particular, the form factors from a single element to all the elements of a
particular patch are treated as a constant scaled by the individual element areas.
Conversely, using the reciprocity principle, the form factors from all elements
of a single patch to any other element are assumed to be identical.
To understand better why this assumption is often valid, it is easiest to use
contrived example in a two-dimensional flatland. In flatland, surfaces are
lines in a plane. Flatland was introduced to radiosity investigation by Heckbert
[122].
Imagine a model consisting of two perpendicular line segments that touch at
one end, as depicted in Figure 7.4(a). In this two-dimensional world the kernel
function G(x, x) (see Chapter 3) of the integral operator depends on the cosine
of the angles between the line connecting the points x and x and the surface
normals at those points, and inversely on the distance, r, between the points
(as opposed to the distance squared in three dimensions). It also depends on
visibility, although this is not an issue in this particular example. Also, aside
from the singularity at the corner where the surfaces meet, G is smooth.
Clearly the value of G is smaller and varies less quickly in cases where x
and xare distant from each other, since the 1/r and cosine terms are then less
sensitive to small changes in x or x. This will be reflected in the values of the
corresponding entries in the matrix K.
The K matrix for this model is shown in Figure 7.4(b).3 The surfaces (line
segments) have been divided into eight elements each. There are a row and a
column corresponding to each element of the two surfaces, resulting in a 16 by
16 matrix. In this particular model, the upper left and lower right quadrants of
the matrix will be identically zero, except for the ones on the diagonal, since
the entries in these quadrants represent interactions between elements on the
same surface. The interactions between elements on two different surfaces are
represented by the upper right and lower left quadrants, enlarged in Figure 7.4(c).
This example has been set up intentionally so that neighboring entries in the
upper right quadrant represent interactions that are closely related in a physical
sense. For example, entries in the upper right-hand corner of the quadrant rep-
resent interactions between elements that are at the far ends of the two surfaces
3
Constant basis functions will be assumed for the discussion of K in the following
sections. However, the ideas are equally applicable to higher order bases.

Radiosity and Realistic Image Synthesis 172

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.3 THE K MATRIX

Figure 7.4: Two flatland surfaces and the upper right quadrant of K. The value
of the entry 2, 10 represents the transport of light from element 10 to element 2.

and that are thus widely separated. According to the behavior of the function G
just described, these entries will be smaller and change less rapidly than those
in the lower left-hand corner of the quadrant, which represent elements that
are closer together. This is demonstrated very clearly in Figure 7.5, which is
a visualization of the upper right quadrant of a 32 by 32 matrix for the same
model.
Returning to the 16 by 16 example, the entries K2,9 and K2,10 , which
neighbor each other on row 2, represent the effect of elements 9 and 10 on

Radiosity and Realistic Image Synthesis 173

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.3 THE K MATRIX

Figure 7.5: Value of entries in the upper right quadrant of the operator K. The
values indicated by the area of the circles.

element 2. If the values of these two entries are almost the same, a single value
could be computed and simply used twice. In essence, a single larger element
composed of elements 9 and 10 would be used as a source of light for element 2.
Why not just eliminate the two smaller elements entirely by joining them
together into a single larger element? This would be reasonable for this par-
ticular interaction. However, it may be that in another row the form factors to
elements 9 and 10 are quite different, and thus it is important to keep the effects
of elements 9 and 10 distinct for those interactions! For example, looking at
Figure 7.5, the difference between the entries of the first two columns is clearly
much more significant in the bottom row than in the top row.
Similarly, two neighboring entries in a single column may be very similar,
indicating that the element corresponding to that column makes a similar con-
tribution to the two receiving elements. In this case, a single larger receiving
element would provide sufficient accuracy. Finally, a whole block of the matrix
may be identical, indicating that the effect of a group of source elements on a
group of receiving elements could be treated as a single interaction.
This discussion has ignored visibility, which may decrease or increase the
coherence between neighboring entries in the matrix. If some polygon fully or
partially obscures the visibility between two other polygons, a series of zeros
will occur in the matrix. Rapidly changing partial visibility may also result in a
portion of the operator that was very coherent becoming less coherent.

Radiosity and Realistic Image Synthesis 174

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.3 THE K MATRIX

In the matrix for a complex three-dimensional model, neighboring entries

are by no means guaranteed to represent neighboring physical situations. Nev-
ertheless, the basic principle remains the same: there will be entries throughout
the matrix that do represent related physical neighborhoods and that can be
grouped along rows, columns, or in blocks to reduce the computation required
to formulate the matrix.
Returning now to the two-level, patch-element hierarchy, it is clear that this
approach has the user make an a priori guess at which entries along rows of
the matrix can be grouped together. This initial grouping is the patch level
subdivision. Elements are only grouped together (into patches) in so far as they
act as source of illumination for other elements. This reduces the n by n matrix
to an m by n matrix of source patches and receiving elements.
The limitations of the two-level algorithm are also clear. First, the grouping
of elements into patches is fixed. The same grouping is applied to every row
of the matrix whether or not the grouped entries would actually be similar in
that row. Thus, even the elements of two surfaces that happen to be quite
close together will be treated as a single patch, although this is not justified.
The ability to group elements into larger or smaller groups depending on the
interaction is important to maintaining accuracy.
Returning to the example of the wall and the table, in a two-level hierarchy
the entire table top might be a patch and act as a source for all elements. This
might satisfactorily account for the effect of the table on the wall, but would be
inadequate to represent the effect of the table on the objects resting on it. A more
detailed representation of the illumination provided by the table is necessary in
this case.
A second limitation is that entries that are similar within a column cannot
be grouped together. This corresponds to the distinction between sources and
receivers in the two-level algorithm. This distinction is justified only in terms
of maintaining a special, highly refined approximation for rendering purposes.
For the purposes of the solution process itself, the inability to group receivers
removes a potential source of efficiency.
Finally, the two-level algorithm requires that the user perform the patch
subdivision. Since the user does not have access to detailed quantitative data
characterizing the interactions being grouped, this is clearly inadequate. An
error metric that can be used to evaluate interactions for possible grouping is
required, as well as an algorithm for automatically constructing the hierarchy.
These limitations are all addressed in the next section, which describes a
general algorithm for constructing multilevel hierarchies of interactions.

Radiosity and Realistic Image Synthesis 175

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

Figure 7.6: Hierarchical quadtrees and interactions.

7.4 Multilevel Hierarchy

Hanrahan et al. [116] have generalized the notion of subdivision hierarchy to
multiple levels, allowing both receivers and sources to be treated at the appropri-
ate level of detail. A basic hierarchical subdivision is diagramed in Figure 7.6,
in which two surfaces have been recursively subdivided into a hierarchy of ele-
ments and groups of elements. The recursive subdivision of each surface result.
in a quadtree of nodes, each of which represents some portion of the surface.
As in the two-level hierarchy, the leaf nodes are the elements, with nodes higher
in the quadtrees representing groups of elements.
In the multilevel hierarchy, energy can be shot from any node to any other
node at any level in the hierarchical subdivision (not only to the leaves as before).
One such interaction is indicated by the arrow in Figure 7.6, where energy from
a group of seven elements (a) is transported to a group of four elements (b).
If energy is shot to a node at a level above the leaves, the nodes below (down
to the leaves) will inherit the energy received above. This operation will be
explained in more detail below. It will be shown that this approach leads to
fewer total shooting (gathering) operations. The key result will be that only O(n)
interactions (form factors) are required to meet a given error tolerance, where n
is the number of leaves of the quadtrees (i.e., the number of elements). Thus,

Radiosity and Realistic Image Synthesis 176

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

while the two-level hierarchy reduced the number of form factors to O(n m),
the general hierarchical algorithm is O(n).

7.4.1 N-Body Problem

Hanrahan et al. [116] gained their inspiration by relating the radiosity problem
to the N-body problem, which addresses the gravitational interactions within a
collection of n particles.
If each of n particles exerts a force on all the other n 1 particles, there
will be O(n2) interactions between pairs of particles to account for. Fast algo-
rithms have been developed [9] by recognizing that the force due to a group of
particles beyond some distance from another particle can be approximated with
a single interaction. Likewise, pairs of such groups, where the whole groups
are separated by some distance, can be considered with a single interaction.
For example, two widely separated galaxies each represent very large groups
of particles, but the gravitational interaction between them can be approximated
by the interaction of two particles, each representing the combined mass of the
stars in the corresponding galaxy. This insight has been refined to develop fast
N-body algorithms by Esselink [79], Barnes and Hut [16], and Greengard and
Rokhlin [106].

7.4.2 Radiosity and the N-Body Problem

The radiosity problem is similar to the N-body problem in that the interactions
(form factors) drop off according to 1/r2. The interactions can also be summed
as with the form factor algebra. On the other hand, the radiosity problem often
starts with large areas to subdivide rather than small particles to group together
(although see the discussion of clustering in section 11.3.5). There is also no
analog to occlusion in the N-body problem (i.e., the gravitational force between
two particles is not affected by intervening particles). Nevertheless, the basic
structure of the hierarchical algorithm remains valid.

7.4.3 Hierarchical Refinement

In contrast to previously discussed algorithms, the multilevel hierarchical algo-
rithm never explicitly builds a matrix. Instead, as the subdivision is performed
recursively, links are built on-the-fly between nodes of the quadtrees associated
with each surface.
A link represents the physical relationship between one set of elements and
another set of elements, which determines the potential for energy transport
between them. The two sets of elements are the leaves of the subtrees below the
nodes at which the links terminate (indicated by the dotted ovals in Figure 7.6).

Radiosity and Realistic Image Synthesis 177

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

In other words, each link will represent a subset of form factors from the original
n by n matrix, that is, the form factors for all pairs of elements in the two
subtrees. A link between a pair of nodes at the lowest level connects two
individual elements and thus represents a single entry in the matrix. Links at
higher levels represent successively larger groups of form factors.
The set of entries in the original n by n matrix, K, that are represented by
a single link will map to regions in the domain of the kernel of the radiosity
integral. Remember that the kernel function is four-dimensional in the case of
a three-dimensional environment, since it is a function of two two-dimensional
points on two surfaces. Thus the local four-dimensional region represented by
a link is the cross product of the two two-dimensional regions represented by
the sets of elements at either end of the link. Since the sets of form factors
encompassed in a single link do not define complete columns of the original
matrix as in the two-level hierarchy, or neat rectangular blocks in the three-
dimensional model, it is more difficult to visualize or to explicitly construct a
matrix representation of the links. Instead, a solution method that relies directly
on the link data structure is used.
To determine the links, one could conceivably build the whole n2 form
factor matrix and then search for subsets of the matrix for which all entries
are similar in magnitude. These would correspond to potential links between
quadtree nodes and could then be grouped together into a single value. This
might, in fact, reduce the solution time for iterative matrix solution methods by
making the matrix-vector multiplications faster, but would not solve the more
important problem of reducing the form factor computation.
Instead, the aim is to develop a method of predicting whether a subset
of form factors will be coherent (e.g., similar in magnitude) before actually
computing the subset of form factors. If the subset is predicted to be coherent,
a single representative form factor can be computed for the subset. Such a
prediction method will be called an oracle in the following discussion. 4
For this discussion of hierarchical algorithms, we will use two data struc-
tures, one for a quadtree node, Quadnode, and one for a link between quadtree
node Linknode (see Figure 7.7). The hierarchical algorithm begins with
a basic recursive procedure called Refine that subdivides the surfaces into
quadtrees and create the links between quadtree nodes (see Figure 7.8). Three
functions are required:

1. Oracle( p, q, ) returns a decision of whether or not to link two quadtree

nodes p and q based on the error that would be incurred if the nodes are
linked rather than linking multiple nodes at lower levels below p and/or q.
If both p and q are already subdivided as far as possible (i.e., their areas
4
The use of an oracle to determine element subdivision is also discussed in [76].

Radiosity and Realistic Image Synthesis 178

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

struct Quadnode {
float Bg; /* gathering radiosity */
float Bs; /* shooting radiosity */
float E; /* emission */
float area;
float ;
struct Quadnode** children; /* pointer to list of four children */
struct Linknode* L; /* first gathering link of node */

};

struct Linknode {
struct Quadnode* q; /* gathering node */
struct Quadnode* p; /* shooting node */
float Fqp; /* form factor from q to p */
struct Linknode* next; /* next gathering link of node q */
};

Figure 7.7: Quadnode and Linknode data structures.

Refine(Quadnode *p, Quadnode *q, float F )

{
Quadnode which, r;
if (Oracle1( p, q, F ) )
Link( p, q );
else {
which = Subdiv( p, q );
if( which == q )
for( each child node r of q ) Refine( p, r, F );
else if (which == p )
for( each child node r of p ) Refine( r, q, F );
else
Link( p, q );
}
}

Figure 7.8: Refine pseudocode.

Radiosity and Realistic Image Synthesis 179

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

SolveSystem()
{
Until Converged {
for ( all surfaces p ) GatherRad( p );
for ( all surfaces p ) PushPullRad( p, 0.0 ); }
}

Figure 7.9: SolveSystem pseudocode.

are below a user-specified minimum area, A ), then the Oracle returns

FALSE.

2. Subdiv( p, q ) is called if the oracle returns true indicating nodes

below p or q should be used. Subdiv returns p or q depending on for
which it appears using lower level nodes will reduce the error. It may
also decide not to choose either node. If the selected quadtree node has
not yet been subdivided then Subdiv performs this operation.

3. Link( p, q ) actually builds the link between p and q by computing the

form factor between the areas represented at nodesp and q of the quadtrees
and stores the form factor in the link data structure.

These three functions and a user-supplied error tolerance, F , provide the

necessary tools to adaptively perform the quadtree subdivision and build the
appropriate links between quadtree nodes. Given two (rectangular or triangular)
surfaces p and q, representing the roots of two quadtrees, the algorithm proceeds
recursively (see the pseudocode in Figure 7.8).
After each ordered pair of surfaces has been handed to Refine, the result is
a network of links that create connections between pairs of quadtree nodes. Note
that every element-to-element pair will be covered. In other words, given two
elements (leaves) i and j, there will be exactly one link from i or is ancestors
to j or js ancestors.
The number of links corresponds to the number of form factor calculations
that must be performed. Using an inexpensive form factor estimate for the oracle,
Hanrahan et al. [116] make a counting argument to show that the number of
links required is O(n) rather than O(n2), where n is the maximum number of
elements if all surfaces are fully subdivided. In fact, this may be a conservative
estimate; their experience has shown that many fewer than n links are usually
created.

Radiosity and Realistic Image Synthesis 180

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

GatherRad( Quadnode *p )
{
1 Quadnode *q; Link *L;
2
3 pBg = 0;
4 for ( each gathering link L of p ) /* gather energy across link */
5 pBg += p * LFpq * LqBs ;
6 for each child node r of p
7 GatherRad( r );
}

Figure 7.10: GatherRad pseudocode.

7.4.4 Solution of the Hierarchical System

Although no explicit matrix has been formed, the set of links created by
Refine defines a linear system that can be solved for the element radiosities.
At each quadtree node, the data structure should contain two radiosity values, Bs
(for shooting) and Bg (for gathering). Information must also be available about
its emission, area and reflectivity. The solution of the system is performed iter-
atively by the function SolveSystem (see Figure 7.9). Each iteration consists
of two steps, performed on each top-level surface (i.e. root of a quadtree):

1. GatherRad gathers energy over each incoming link, converting Bs at

one end into Bg at the other (line 5 of Figure 7.10).

2. PushPullRad, pushes the received energy Bg down to the children of

each quadtree node, and pulls the results back up the quadtrees by area
averaging (see Figure 7.11), thus preparing the Bs for the next iteration.
On the way recursively down the quadtree, the gathered radiosity of each
node is simply added to the nodes below (line 9 of Figure 7.11). The
radiosity of an area is not diminished by cutting the area into parts since
radiosity has units of power per area, thus the simple summing of radiosity
values.

At the leaves of the quadtree, the sum of the gathered radiosity pushed
down the tree is added to the leaf nodes own gathered radiosity and
emitted radiosity and the result transferred to the leaf node's shooting
radiosity (line 3 of Figure 7.11). This quantity is returned up the tree
(line 14). Since radiosity is in units of power/area, the radiosity pulled up
the tree is the average over a node's children (line 10). For example, if

Radiosity and Realistic Image Synthesis 181

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

PushPullRad( Quadnode *p, float Bdown )

{
1 float Bup , Btmp ;
2 if ( pchildren == NULL ) /* p is a leaf */
3 Bup = pE + pBg + Bdown ;
4 else
5 {
6 Bup = 0;
7 for (each child node r of p)
8 {
9 Btmp = PushPullRad(r, pBg + Bdown );
10 Bup += Btmp * rarea
parea
11 }
12 }
13 pBs = Bup;
14 return Bup;
}

Figure 7.11: PushPullRad pseudocode.

four children all have the same radiosity, their parent node will have the
same radiosity, since radiosity is measured in power/area. (By contrast,
the power for the parent is the sum of the powers of the children).

The two steps in SolveSystem are performed iteratively until less than a
user-specified maximum change occurs in the radiosity values from one iteration
to the next. Thus, it may be advisable to store a copy of Bs before zeroing it
out and then run a comparison of the copy and new shooting radiosities. This
solution process corresponds to the iterative relaxation methods described in
Chapter 5.

7.4.5 The Oracle Function

The Oracle function plays a key role in the efficiency of the overall hi-
erarchical algorithm. The job of the oracle is to estimate the error that will be
introduced by linking two quadtree nodes, rather than creating a series of links
at lower levels. In essence, the oracle must answer the equivalent question:
By linking nodes p and q, what is the error introduced if a constant5 value is
5
In fact, the constancy test is the result of the use of constant hierarchical basis
functions. This will be generalized in Part II of this chapter.

Radiosity and Realistic Image Synthesis 182

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

Boolean Oracle1( Quadnode *p, Quadnode *q, float Fe )

{
if ( parea < Ae and qarea < Ae )
return( FALSE );
if ( EstimateFormFactor( p, q ) < Fe )
return( FALSE );
else
return( TRUE );
}

Figure 7.12: Oracle1 pseudocode.

assigned to all Kij entries in the matrix, where i is a descendant of quadtree

node p and j is a descendant of q? It is not immediately clear how to answer
the above question. For example, should the error norm reflect the estimated
error in the entries of the matrix (i.e., the form factors) or the error in the final
radiosity function? In addition, since the goal is to eliminate evaluations of the
form factor, the oracle must be less expensive to evaluate than the form factor.

The Oracle as Form Factor Estimate

In the work reported by Hanrahan et al., the value of the oracle is derived from
an estimate of the unoccluded form factor (see Figure 7.12):
cos
F pq q (7.4)

The estimate is computed at the centers of the areas represented by quadtree
nodes p and q. The factor q is the solid angle subtended by a sphere (or disk)
surrounding the area of node q (see Figure 7.13).
This provides an upper bound on the error of any single entry in K. The
rationale is that the form factor estimated by equation 7.4 can only get smaller if
the real area q is used, or if there is occlusion, or if a descendant of q with smaller
area is selected. In [116] both Fpq and Fqp are estimated. If either estimate is
larger than a given F, the element corresponding to the larger is subdivided.
When both are below the threshold, a bidirectional link is established. This
oracle is efficient and simple to implement, and is reported to work well in
experiments by Hanrahan et al.

Radiosity and Realistic Image Synthesis 183

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

Figure 7.13: Geometry for simple oracle function that estimates unoccluded form
factor.

Other Oracle Possibilities

The oracle described above establishes a threshold for individual errors in the
K matrix. Other norms are possible and may offer a better prediction of error.
The assumption implicit in creating a link is that the set of entries in the full
matrix represented by the link are approximately equal in magnitude. The above
test measures the maximum possible form factor value (and thus bounds the
individual errors in K). However, a better test might be to estimate the potential
variability in the differential form factor. This would require either an analytic
formula for the derivative of the form factor or a multipoint quadrature and the
use of finite differences across the two areas at either end of the link. Although
this test would be more expensive, it might also lead to less subdivision and
fewer total links. In fact, Hanrahan et al. perform a variation of this scheme
when they take into account knowledge about partial visibility between surfaces
in the adaptive algorithm discussed in the next section.

7.4.6 Progressive Refinement of the Hierarchy

The hierarchical algorithm outlined above makes a priori decisions about the
level of the hierarchy at which to create the links, then solves the resulting sys-
tem. Thus, the oracle makes decisions based solely on the geometry of the form
factor estimates, independently of how much energy will eventually be trans-
ferred across individual links. An adaptive a posteriori version (see Figure 7. 14)

Radiosity and Realistic Image Synthesis 184

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

HierarchicalRad(float BF )
{
Quadnode *p, *q;
Link *L;
int Done = FALSE;
for ( all surfaces p ) pBs = pE;
for ( each pair of surfaces p, q )
Refine( p, q, BF );
while ( not Done ) {
Done = TRUE;
SolveSystem(); /* as in Figure 7.9 */
for ( all links L )
/* RefineLink returns FALSE if any subdivision occurs */
if( RefineLink( L, BF ) == FALSE )
Done = FALSE;
}
}

Figure 7.14: HierarchicalRad pseudocode.

of the hierarchical algorithm can make better decisions about the subdivision,
thus leading to more efficient performance. The a posteriori algorithm estab-
lishes a threshold for the oracle based on the amount of energy transferred across
any individual link. The threshold BF is based on the radiosity, the element
area, and the form factor, or B F A. Since the radiosities, B, are not known
a priori, the algorithm proceeds adaptively (see Figures 7.14 and 7.15), using a
modified Oracle2 function (see Figure 7.16).
Using the modified oracle, links are formed only from quadtree nodes from
which a significant amount of energy is ready to be shot. In the first pass,
links are formed only at the highest level unless the shooting surface is a light
source. As the algorithm progresses and more surfaces receive light to reflect,
old links are broken and new links formed at lower levels as necessary. The
link structure converges when all links carry approximately the same amount of
energy.
An additional enhancement can be made by storing visibility information
with the links. When the link is created, the form factor is computed and
information describing the visibility between the two elements is recorded (i.e.,
whether p and q are fully visible, fully occluded, or partially visible to each

Radiosity and Realistic Image Synthesis 185

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.4 MULTILEVEL HIERARCHY

int RefineLink(Linknode *L, float BF )

{
int no_subdivision = TRUE;
Quadnode* p = Lp ; /* shooter */
Quadnode* q = Lq ; /* receiver */

if ( Oracle2(L, BF ) {
no_subdivision = FALSE ;
which = Subdiv( p, q );
DeleteLink( L );
if ( which == q )
for (each child node r of q) Link( p, r );
else
for (each child node r of p) Link( r, q );
}
return(no_subdivision);
}

Figure 7.15: RefineLink pseudocode.

float Oracle2( Linknode *L, float BF )

{
1 Quadnode* p = Lp ; /* shooter */
2 Quadnode* q = Lq ; /* receiver */
3 if ( parea < A and qarea < A )
4 return( FALSE );
5 if ( pBs == 0.0 )
6 return( FALSE );
7 if (pBs * pArea * LFpq ) < BF );
8 return( FALSE );
9 else 10 return( TRUE );
}

Figure 7.16: Oracle2 pseudocode.

Radiosity and Realistic Image Synthesis 186

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.5 HIERARCHICAL BASIS FUNCTIONS

Polygons 98
Potential elements 175964
Potential interactions 15481576666

Quadtree Nodes 5674

Elements 4280
Interactions 11800
Totally-invisible 4605 39.0%
Totally-visible 4519 38.3%
Partially-visible 2676 22.7%
Tests
Refinement tests 14149
Totally-invisible refines 3901 27.6%
Pre-Totally-invisible refines 0 0.0%
Totally-visible refines 5414 38.3%
Pre-Totally-visible refines 4128 29.2%
Partially-visible refines 4834 34.2%
Partial visibility tests 10021
Ray tests 53187
Visibility tests 3545
Ray tests 56720

Table 7.1: Statistics for color plates 18-22 (after Hanrahan et al.).

other). Clearly, the fully occluded case will not benefit from further subdivision.
The fully visible case can have a more relaxed threshold set than the partially
visible case, for which one might expect larger fluctuations in the form factor
kernel across the portions of the surfaces represented by p and q.

7.4.7 Experimental Results

Reports of experiments [116] using the HierarchicalRad algorithm are very
encouraging (see Table 7.1). Color plates 1821 show the links formed during
the progressive refinement process. The multiple images are provided in order to
show the links at the various levels of the hierarchy. Links are also color-coded
by visibility (dark blue: fully occluded, white: fully visible, green: partially
occluded). Color plate 22 shows a final image with texture mapping added (see
Chapter 10 for a discussion of texture mapping). The statistics in Table 7.1
provide numerical evidence of the algorithms ability to produce an image with
a very limited number of interactions, (e.g., only about 12 thousand out of a
potential 15 billion interactions are created).

II. Hierarchical Basis Functions and Wavelets

7.5 Hierarchical Basis Functions

The hierarchical methods just described can alternatively be characterized in
terms of hierarchical basis functions. This view also provides a framework for

Radiosity and Realistic Image Synthesis 187

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.5 HIERARCHICAL BASIS FUNCTIONS

Figure 7.17: Hierarchical basis set used by Hanrahan et al.

developing other hierarchical basis functions as well as some understanding of

the properties that lead to efficient algorithms. The basis functions discussed
in previous chapters all have the property that their support is limited to the
elements adjacent to the node at which the basis function is situated. Hierarchical
basis functions will be developed by relaxing this restriction.
As an example of a hierarchical basis set, Figure 7.17 shows a binary tree of
hierarchical box bases in one dimension. The basis set can be constructed either
top-down or bottom-up. In the top-down approach, a single box basis, NL,1, at
the top level, L, is subdivided in half, producing two basis functions, NL1,1
and NL1,2, with the same combined support in the domain. These are in turn
subdivided recursively until a preset minimum level is reached. The bottom-up
approach simply reverses this process, beginning with the lowest level of box
bases, N1,1 to N1,n, and recursively grouping two bases into a new, wider basis
one level up. In the radiosity context, a binary tree (in 2D) or quadtree (in 3D)
of bases would be constructed for each surface.
Figure 7.18 shows a one dimensional function approximated by a set of basis
functions at each level of the hierarchy. The coefficients of each basis function
represent the average radiosity over the support of the basis. These coefficients
would be, for example, the By terms after applying the PushPullRad function
in Figure 7.11.

Radiosity and Realistic Image Synthesis 188

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.5 HIERARCHICAL BASIS FUNCTIONS

Figure 7.18: Function represented at each level of hierarchy.

The hierarchical basis set is redundant in the sense that any function repre-
sentable at one level of the hierarchy can also be represented one level below
(with two coefficients having values identical to the higher level coefficient).
The reverse is not the case, however. Thus, the use of hierarchical systems does
not lead to an increase in the size of the function space spanned by the bases. In
particular, the multilevel hierarchical system described in the previous sections
spans the same piecewise constant space of functions as a nonhierarchical set of
box functions consisting of only the lowest level of the hierarchy. Thus, the use
of a hierarchical system itself cannot produce an approximation B with a better
fit to the actual function B. Instead, the use of hierarchy leads to sparser linear
operators (i.e., fewer form factors) and thus more efficient algorithms.
The hierarchical set of basis functions, being redundant, would at first appear
to be less efficient since there are more bases to represent the same space of
functions. The hierarchical system, however, gains its efficiency from the fact

Radiosity and Realistic Image Synthesis 189

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

that, unlike the nonhierarchical formulations, not all pairs of basis functions
require interaction terms (form factors). For example, using the subscripting
shown in Figure 7.17 and a superscript to represent the surface at the root of
p
the tree, N3,4 would be the fourth basis function on the third level, for the L
level quadtree associated with surface p. In the nonhierarchical formulations K
has entries for every pair (1, j) of elements (represented as the lowest leaves in
the hierarchical system, with the basis functions at level 1). In contrast, in a
hierarchical system, if a link is established between nodes of the quadtrees on
p q
surfaces p and q, for example, between N3,4 and N5,1 , then no more links can
p q
be made from N3,4 (or any of its children on levels 1 and 2) to N5,1 (or any of
its children on levels 4 and below). Thus, this single link represents all pairings
p q
of leaves in the quadtree below N3,4 and N5,1 . In total, the single form factor on
p q
the link from N3,4 to N5,1 will represent 2 (31)
2(51) = 64 form factors at the
lowest level of the binary trees in two dimensions, and 4(31) 4(51) = 4096
form factors for the quadtrees in three dimensions!

7.6 Wavelets
A growing body of theory and applications is associated with the wavelet family
of bases [27, 185]. It is beyond the scope of this book to provide a full discussion
of wavelet theory. However, a short introduction will make it possible to cast the
hierarchical radiosity formulation in terms of a simple example from the family
of wavelets. The intention is to introduce some of the basic ideas necessary to
extending hierarchical methods to new basis sets. These new basis sets have
properties that allow them to better approximate the radiosity function and may
possibly provide even sparser representations of the integral operator. Details
can be found in [102].

7.6.1 Haar Basis

Building a wavelet basis begins with two functions, (x) (called the smooth
function), and (x) (called the detail function). An example from a simple
wavelet basis known as the Haar basis is shown in the center of Figure 7.19. By
examining the two box basis functions, N1 and N2, and the smooth and detail
Haar functions, it can be seen that the two sets both span the same space of
functions. In other words, a linear combination of either the two boxes or the
two Haar bases can represent any piecewise constant function, F(x), over the
two intervals. An example is shown at the bottom of the figure. The coefficient
of the function will represent the average over its nonzero range, and the
coefficient of the function represents the difference from the average.

Radiosity and Realistic Image Synthesis 190

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.19: Two consecutive box basis functions are equivalent to the and
functions of the Haar basis.

Given the above observation, a set of n (where n is a power of 2) box basis

functions can be replaced with a hierarchy constructed from the Haar and
functions (see Figure 7.20) using a bottom-up approach. Beginning with eight
box basis functions, we can replace these by four sets of and functions as
shown in the first row of Figure 7.20. The resulting four and four functions
are then reordered, grouping each set together. In the second row it can be seen
that the four functions are exactly the same in structure as the original box
basis functions, but twice as wide. Thus, these four bases can be rewritten
as two and two functions, now twice the width. Once again these can
be reordered. Finally, in the bottom row the two new wider functions are
rewritten as one and one function. This results in the Haar wavelet basis
consisting of one smooth L,1 basis (where L, the top level, equals log2(n) for
n box bases) at the top level and a pyramid of detail basis functions (see
Figure 7.21). Just as any piecewise constant (over unit intervals) function F(x)
could be represented by a linear sum of the original box basis functions:
n
F(x) = f N (x)
i=1
i i (7.5)

Radiosity and Realistic Image Synthesis 191

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

wavelet basis
Figure 7.20: Construction of the hierarchical Haar basis.

F(x) can also be represented by a linear combination of the Haar wavelet basis
functions:
L n/2(i 1)
F(x) = 1,L L,1 (x) +
i=1 j=1
i, j i, j (x) (7.6)

The coefficients of the box basis functions, fi , represent the local value
of F(x). In the case of the wavelet basis, the coefficient 1,L represents the
average of the function overall, and the i,j represent the local detail or variation
away from the average in F(x) at each level i. This immediately points to the
key advantage of the wavelet basis. If the function being represented is locally
smooth (in this case constant) over a portion of its domain, then the coefficients
representing the local detail will be zero (note the third coefficient in the L 2

Radiosity and Realistic Image Synthesis 192

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.21: Haar wavelet basis.

row in Figure 7.22). While the coefficients of the original hierarchical box
functions represent local averages, the coefficients of the Haar wavelet bases
represent local differences and thus are zero in locally constant regions.
The projection of a one-dimensional function into the Haar basis can be
thought of as passing a high-pass ( or detail) and low-pass ( or smooth)
filter over the function. The filters are then applied again recursively on the
coefficients obtained in the low-pass filtering operation.
Referring again to the piecewise constant function in Figure 7.22, low and
high pass filters are implemented by taking local averages and local differences,
respectively. One application of low-pass filtering gives the coefficients 11.5,
10.5, 8.0, and 10.0. High-pass filtering gives the four coefficients, 0.5, 0.5, 0.0
and 1.0, shown in the row labeled L 2, with the coefficient 0.5, for example,
being the difference between the local average, 11.5, and the two segments of the
original function, 11.0 and 12.0. The coefficient 0.0 corresponds to a flat region
in the function (i.e., no detail). The process then recurses on the coefficients
obtained by low-pass filtering, leading to two average values 11.0 and 9.0, and
differences of 0.5 and 1.0. Finally, the last filtering pass results in the overall
average value of 10.0, which becomes the coefficient of the highest level smooth
basis, and a coefficient of 1.0 for the highest level detail basis. The Haar basis
thus results in 7 coefficients and one coefficient. The value of the function
F(x) as represented by the basis functions can then be obtained by evaluating
equation 7.6 with these coefficients.

Radiosity and Realistic Image Synthesis 193

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.22: Piecewise constant function represented with the Haar basis (top)
and the hierarchical box basis (bottom).

7.6.2 Vanishing Moments

The purpose of developing the preceding alternate basis for representing piece-
wise constant functions is to understand how it leads to a sparse representation of
a function. In particular, the goal is to find a sparse representation of the radios-
ity kernel function, and similarly a large number of zeros in the discrete operator,
K. Wavelets are finding many uses in data, image, and signal compression for
similar reasons.
If it is possible to find a basis set that leads to a sparse representation, and
if one can predict in advance where the nonzero regions (entries in K) will be,
then it will be possible to reduce greatly the number of form factors required.
Concepts presented below in one dimension will be extrapolated to apply to the
multidimensional kernel of the radiosity problem.
The key concept is that of vanishing moments. A function (x) is said to

Radiosity and Realistic Image Synthesis 194

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.23: Basis with two vanishing moments in detail functions.

have M vanishing moments if,

i
(x) x dx = 0 for i = 0, ... , M 1 (7.7)

For example, a function has one vanishing moment if its integral is zero, since
x0 is a constant. A function has two vanishing moments if when integrated
against any linear function the result is zero, and so on for quadratic, cubic, and
higher order polynomials. The Haar wavelet detail function has one vanishing
moment since any constant function when integrated against it will be zero (i.e.,
vanish). For example, in Figure 7.22 the third coefficient on the L 2 row is
zero since the local region of the function is flat.
Figure 7.23 depicts a set of four bases, two smooth and two detail, that
also span the same piecewise constant function space [102], but in this case, the
detail functions have two vanishing moments (call this the ^2 basis). In other
words, any linear function will vanish when integrated against these functions.
The price that one must typically pay for more vanishing moments is a wider
support, in this case over four intervals rather than the two of the Haar basis.
Basis sets can be constructed with varying numbers of vanishing moments and
abilities to span higher order function spaces, with associated costs in terms
of their support and the difficulty of evaluating integrals in which they appear.
Construction of such bases is not pursued here. The reader in encouraged to
investigate the expanding literature on wavelets for this purpose.

7.6.3 Vanishing Moments and Sparse Representations

How do vanishing moments in the basis functions lead to sparse representations
of functions and integral operators? The answer can be seen in the example of

Radiosity and Realistic Image Synthesis 195

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.24: Piecewise constant function represented by 9 hierarchical Haar

basis functions versus 16 box bases.

Figure 7.24. In this specific example, the piecewise constant function can be
exactly represented as a weighted sum of 16 individual box bases. Alternatively,
it can be represented by a weighted sum of only eight hierarchical Haar detail
basis functions plus the average value of the overall function represented by the
weight of the single smooth basis function. For areas in which the function is
flat (i.e., constant), the projection6 onto lower-level detail functions vanishes,
and thus these basis functions are not necessary. Similarly, projecting onto bases
with two vanishing moments will result in zero weights when the function is
linear across the support of the basis. In general, functions will not exhibit
exactly constant or linear behavior across the support of a basis. However,
regions that are almost constant (or linear) will result in very small coefficients.
In this case, the coefficients (weights) can be set to zero with only a small error
in the approximation.
The hierarchical representation of a 2D matrix using the Haar basis is some-
what similar to the hierarchical mipmap representation of a texture map [266].
6
The projection (more precisely the orthogonal projection) of one function onto
basis described in Chapter 3. The projection involves finding the coefficient or weight
of each basis function so as to minimize the norm of the difference between the original
function and the sum of the weighted basis functions.

Radiosity and Realistic Image Synthesis 196

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.25: The two-dimensional nonstandard pyramid algorithm (from Gortler

et al., 1993). (a) The full n by n matrix. (b) A horizontal high-pass and low-
pass filter on each row results in n/2 smooth and n/2 detail coefficients. (c) A
vertical pass on this results in n/4 detail-detail coefficients, n/4 detail-smooth
coefficients, n/4 smooth-detail, and n/4 smooth-smooth coefficients. (d) The
process is recursively applied to the smooth-smooth quadrant.

To construct a mipmap, a two-dimensional low-pass filter 7 is passed over the

original image, giving an image of 1/4 the size. The same filter is applied again
on the result recursively, resulting in a pyramid of images. Each successive
image moving up the pyramid averages the values from below and thus contains
less detail; the final image consists of a single value equal to the average value
of the original image.
The construction of a representation of a 2D matrix (or image) using the
Haar basis begins with the application of low-pass and high-pass filters along
the rows and columns of the matrix (see Figure 7.25). For the Haar basis,
low-pass filtering consists of averaging together pairs of entries along the row
or column and high-pass filtering of taking differences between entries. Each
filter is first applied to the matrix in the horizontal direction (along the rows).
If there are n entries in a row, the result will be n/2 smooth coefficients and
n/2 detail coefficients. Following the filtering pass, the coefficients resulting
from each filter are grouped together (see letter (b) of Figure 7.25), forming a n
by n matrix for which the first n/2 entries in each row are smooth coefficients,
7
A 2D box filter is equivalent to the tensor product of the one-dimensional smooth
Haar basis F.

Radiosity and Realistic Image Synthesis 197

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

and the last n/2 entries are detail coefficients. The filters are then applied in
the vertical direction to this matrix and the results are similarly reorganized to
give an n, by n matrix for which the coefficients in the four quadrants consist of
the detail-detail, detail-smooth, smooth-detail, and the smooth-smooth filtering
results (see letter (c) of Figure 7.25). This filtering and reorganization process
is then repeated recursively on the upper right (smooth-smooth) quadrant.
The final result of this decomposition is a hierarchy of representations of the
2D matrix (in what has been termed a nonstandard two-dimensional Haar basis
[26]). The differences between this representation and a mipmap are instructive.
The levels of a mipmap contain low frequencies (averages) in all levels of
the image. The Haar basis also begins at the top with a single average value;
however, at each level moving down from the top level the coefficients represent
the variation of the matrix values from the representation of the previous level.
The possibility for efficiency lies in taking advantage of the fact that smooth
regions of the matrix result in near zero valued coefficients that can be ignored.
The results of applying the above decomposition process to the K matrix
are shown in Figures 7.26 and 7.27 for the same flatland example (two per-
pendicular line segments) used in Figure 7.4. The top box in Figure 7.26 shows
the 32 by 32 upper right quadrant of a 64 by 64 matrix, with the dot size pro-
portional to the form factor in the original matrix. In the two boxes below,
the matrix has been projected into the nonstandard basis (examples are shown
for the Haar and ^2 basis). The three quadrants in the lower left of the boxes
are the detail-detail, detail-smooth and smooth-detail quadrants following one
complete filtering pass (see letter (c) of Figure 7.25). The smooth-smooth quad-
rant is not shown. Instead, the next three quadrants moving diagonally toward
the upper right show the result of the second complete filtering pass applied
the missing smooth-smooth quadrant. The recursive application of this process
generates the progressively lower detail representations of the matrix, as shown.
The four boxes at the bottom of Figure 7.26 show the error in the reconstruction
of the matrix (as compared to the original), after discarding (setting to zero) all
but the 32 or 64 largest terms.
Figure 7.27 shows a similar example for the case of two parallel line seg-
ments. The important thing to note is that only 32 or 64 entries, compared to the
original 32 32 = 1024 matrix coefficients have been used to reconstruct the
original matrix while incurring only very small errors. Also, note the reduced
error when applying bases with two vanishing moments (the ^2 basis).

7.6.4 A Wavelet Radiosity Algorithm

A full radiosity application differs in two major ways from the above examples.
First, the radiosity domain consists of two-dimensional surfaces embedded in

Radiosity and Realistic Image Synthesis 198

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.26: Projecting a flatland kernel (from Gortler et al., 1993). The original
matrix shows the kernel function between two perpendicular line segments that
meet in a corner discretized into a 32 by 32 grid. Darker entries represent larger
kernel values. The kernel values are greatest in the lower left corner, where the
two segments meet and 1/r goes to infinity. This kernel is projected into both the
nonstandard two-dimensional Haar basis and the nonstandard two-dimensional
basis with two vanishing moments. In both of these representations many of
the 1024 coefficients are small. In both cases, only 32 and 64 of the largest
coefficients have been selected to reconstruct an approximate kernel. An error
matrix (the difference between actual kernel and approximate kernel) is shown.
This demonstrates that a low error can be achieved with very few entries from
the projection.

Radiosity and Realistic Image Synthesis 199

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.6 WAVELETS

Figure 7.27: This figure shows the application of the wavelet projection to the
flatland configuration of two parallel line segments. The kernel is largest along
its diagonal, where points on the two segments lie directly across from each
other. Note the greater sparsity provided by the projection into the basis with
two vanishing moments.

three-dimensional space. Thus, the kernel function is four-dimensional, since

each basis function is two-dimensional. This does not present any important
conceptual problems or changes in the algorithm, but makes illustration in the
fashion of Figures 7.26 and 7.27 difficult.
The more important difference is that the goal is not to compute the full
matrix and then decompose it to find the significant terms, since this entails
computing n2 form factors, where n is the number of elements. Instead, as in

Radiosity and Realistic Image Synthesis 200

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.7 IMPORTANCE MESHING

the hierarchical algorithms outlined in the previous section, an oracle is relied

upon to predict which terms will be significant. The numerical integration of
the kernel (form factor computation) is then only performed for the associated
pairs of bases that are determined to be significant by the oracle.
The full family of wavelet bases provide options for continuity (e.g., linear,
quadratic, and higher order bases) beyond the piecewise constant bases discussed
here. They also can exhibit more vanishing moments leading to sparser linear
operators. In general, more vanishing moments will require a wider support for
the basis functions and higher order continuity leads to higher costs in evaluating
the form factor integrals and developing appropriate oracle functions. Further
study is required to assess these tradeoffs to develop optimal hierarchical bases
for the radiosity application.
Although the section above cannot provide a complete description of the
algorithms required to project the integral operator onto the wavelet bases, it is
hoped that some understanding of the potential of hierarchical bases has been
provided. The reader is directed to the growing body of literature on wavelets
and their applications to the solution of integral equations for a detailed study
of this new research topic [27, 185].

III. Importance-Based Radiosity

7.7 Importance Meshing

All of the discussion above and in the previous chapters has assumed the goal of
generating a view-independent solution. Consequently, the error metrics we have
described assume all surfaces to be equally important. What if, in contrast, the
goal is to create only a single image of the environment from a given viewpoint,
or multiple images confined to a small area of an extensive model? This is
a common situation in architectural rendering, where models may consist of
buildings containing many rooms and floors.
It would be preferable in this case to concentrate the effort on closely ap-
proximating the radiosity of surfaces visible in the image. It is still necessary to
consider all surfaces, however, since surfaces that are not visible may contribute
indirectly to the illumination of visible surfaces.
Making decisions about where and when to subdivide surfaces given the
knowledge of a fixed viewpoint for the final image requires an error metric that
takes into account the relative importance of a given surface to the image. Devel-
oping such a view-dependent error metric will lead to more efficient algorithms
for this case.

Radiosity and Realistic Image Synthesis 201

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.7 IMPORTANCE METHODS

Figure 7.28: Just as light emission E interreflects about an environment resulting

in a radiosity at the surfaces, a receiver function R emanating from the eye
interrflects among surfaces resulting in the surfaces having varying values of
importance to the final image.

7.7.1 The Importance Equation

Smits et al. [220] have developed an importance-driven approach to radiosity
that uses the above ideas. Illumination models such as the radiosity equation
capture the transport of photons from the light source to the eye. Ray tracing
from the eye typifies a dual process in which rays representing photon paths are
traced in the reverse direction from that taken by the light. Similarly, just as the
radiosity problem can be stated as

K B(x) = E(x) (7.8)

where K is a linear operator on the unknown radiosity, B, and E is a source

term, one can write an adjoint equation for the dual process:

K* Y(x) = R(x) (7.9)

In this equation K* is again a linear operator, but this time it acts on the unknown
importance Y(x). R(x) is a receiver function. The adjoint operator K* in this
case is simply KT. In neutron transport theory, a similar adjoint to the transport
equation is developed to increase the efficiency of solutions for the flux arriving
at a small receiver [68, 152].
Intuitively, one can think of importance as a scalar function over the surfaces,
just like radiosity. The value of Y(x) represents the importance of point x to

Radiosity and Realistic Image Synthesis 202

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.7 IMPORTANCE METHODS

Figure 7.29: Geometry for initializing receiver function

the final image, given a receiver (or eye) position. The receiver function R(x)
acts much like the light sources in the radiosity equation. It can be thought of as
resulting from a spotlight at the eye location emanating importance in the view
direction.
The receiver function can be initialized by assigning each point a value
related to the points visible projected size in the image; thus areas not seen in
the image have a zero receiver value. More formally,

cos x
if x is visible in the image cos x
R(x) = r 2 = V(eye, x)
cos eye r 2
0 otherwise
(7.10)
where the terms are shown in Figure 7.29. In the discrete function, the receiver
value Ri of each element i is proportional to the area of the image covered by
the element. If the eye point moves and multiple images are to be rendered from
the solution, the receiver function might be nonzero at any point ever visible in
the sequence
Just as light interreflects about an environment, the importance from the eye
interreflects among surfaces (see Figure 7.28). Clearly, a point that is visible in
the image will have a large importance, but a point on another surface that has
a large potential to reflect light to a visible surface will also have a significant
importance value.

Radiosity and Realistic Image Synthesis 203

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.7 IMPORTANCE METHODS

7.7.2 Importance-Based Error

If the approximate linear operator and radiosity are given by K and B , respec-
tively, and the error K in the operator is given by

K = K K (7.11)

then a view-dependent error norm is provided by

Y(x) K B (x) (7.12)

A detailed derivation call be found in [220]. The importance function Y(x)

is unknown, just like the radiosity function B(x). Thus, the best that can be
hoped for is to compute an approximation Y in a manner similar to that used to
approximate the radiosity function. A more useful importance-based error norm
is thus provided by
(x) K B (x)
Y (7.13)

When evaluated at a point, this norm gives a measure of the error approximation,
weighted by the importance of the point to the image and by the magnitude of
the radiosity at that point.
This leads to an algorithm (see Figure 7.30) in which one simultaneously
solves for the approximate radiosity and importance functions, making adap-
tive subdivision choices based on the importance-based metric of equation 7.13.
Clearly, minimizing the error in either the radiosity function or the importance
function alone will reduce the overall error. Importance based algorithms gain
their efficiency by taking into account both radiosity and importance, concen-
trating more effort on the radiosity function in regions of high importance, and
more effort on the importance function in brightly lit regions (i.e., high radiosity
values).
Solving for importance is very similar to solving for radiosity. If the radios-
ity associated with element i is
n
Bi = Ei + i B F
j=1
j ij (7.14)

then the importance of element i is given by

n

Yi = Ri +
j=1
rj Yj Fji (7.15)

Note the switch in the indices of and F.

Radiosity and Realistic Image Synthesis 204

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.8 HIERARCHICAL RADIOSITY AND IMPORTANCE

ImportanceDrivenRad( float F )
{
float eps;
for ( all surfaces p )
{
pBs = pE;
pYs = pR; /* visible area in image */
for (all mutually visible pairs of surfaces (q,q))
{ Link(p, q); Link(q, p); }
}
/* beginning with large tolerance, reduce it after each iteration */
for ( initially large diminishing to a small error threshold )
{
SolveDualSystem();
for (each link L)
RefineLink(L, ); /* see Figure 7.15 */
}
}

Figure 7.30: ImportanceDrivenRad pseudocode.

Color plates 23 and 24 show independent radiosity and importance solutions,

respectively. The more interesting image is color plate 25, which shows the
combined radiosity and importance solution. The yellow areas with a high
combined radiosity and importance are the most critical to the final accuracy of
the image.

7.8 Hierarchical Radiosity and Importance

7.8.1 Pseudocode
The importance-based radiosity approach solves simultaneously for two un-
known functions, the radiosity function B(x) and the importance Y(x) (see
Figure 7.31). The discrete forms of the two equations after projection onto
constant basis functions are B = KE and Y = KT R. The error norm to be
minimized is then based on

YT KB (7.16)

Radiosity and Realistic Image Synthesis 205

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.8 HIERARCHICAL RADIOSITY AND IMPORTANCE

SolveDualSystem()
{
Until Converged
{
for (all surfaces p) GatherRadShootImp(p);
for (all surfaces p) PushPullRad(p, 0.0);
for (all surfaces p) PushPullImp(p, 0.0);
}
}

Figure 7.31: SolveDualSystem pseudocode.

GatherRadShootImp( Quadnode *p )
{
pBg = 0;
for (each gathering link L of p)
{
/* gather energy across link */
pBg += pp * LF * LqBs ;
/* shoot importance across link */
pYg += pp * LF * LqYs ;
}
for (each child q of p) GatherRadShootImp( q );
}

Figure 7.32: GatherRadShootImp pseudocode.

where K is the matrix of errors in the operator K. Pseudocode for the hierar-
chical importance-based algorithm is provided in Figures 7.31, 7.32, and 7.33.
The input to the algorithm consists, as before, of the geometry and reflectance
properties of the surfaces, the light source emissions, and additionally, the initial
importance R of each surface. This is typically the visible area of the surface in
the desired image (or images). The algorithm then recursively solves the dual
system, subdividing elements and creating new links until all interactions are
within given tolerances for the product of radiosity and importance.
The changes from the basic hierarchical code (described in section 7.4.3)

Radiosity and Realistic Image Synthesis 206

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.8 HIERARCHICAL RADIOSITY AND IMPORTANCE

PushPullImp( Quadnode *p, float Ydown )

{
1 float Yup , Ytmp ;
2 If ( pne == NULL) /* p is a leaf */
3 Yup = pR + pYg + Ydown;
4 else
5 {
6 Yup = 0
7 for (each child node r of p)
8 {
9 Ytmp = (pY + Ydown) * rarea
parea
10 Yup += PushPullImp(r, Ytmp );
11 }
12 }
13 pYs = Yup;
14 return Yup;
}

Figure 7.33: PushPullImp pseudocode.

are minimal and much of the pseudocode is reused. The following

must be made.

Fields to hold the receiver value R, and the importance Y must be added
to the Quadnode structure.

Importance is shot at the same time that energy is gathered over the links
(see Figure 7.32).

Pushing and Pulling importance is similar to the same operations for ra-
diosity except that the area averaging is reversed (see Figure 7.33). This
is due to the fact that radiosity is energy per unit area. Thus, moving up
the quadtree is an averaging operation. Importance, in contrast, is pro-
portional to area. Thus, moving one level up the tree requires summing
importance (line 10 in PushPullImp). Transferring radiosity and im-
portance down one level is just the reverse; radiosity is transferred directly
while importance must be parceled out according to the area of the chil-
dren (line 9) (compare PushPullImp) in Figure 7.33 to PushPullRad
in Figure 7.11).

Radiosity and Realistic Image Synthesis 207

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 7. HIERARCHICAL METHODS
7.8 HIERARCHICAL RADIOSITY AND IMPORTANCE

The oracle function must be modified to include multiplying by the

current element importance Y as well as the form factor on the link and
the radiosity and area of the element (line 7 in Figure 7.16).

7.8.2 Example Results

Color plates 26-31 show a series of images of the radiosity-importance solution
(color plates 26, 28, 30) and the importance-only solution (color plates 27, 29,
31) for a mazelike environment [220]. As the eye point is backed away from
the original point from which importance was propagated, it is apparent that the
algorithm has concentrated effort on regions that are either visible in the original
image or that contribute significantly to the light eventually arriving at the eye.
The timings given by Smits et al. indicate a performance increase of two to three
orders of magnitude for this environment.

Radiosity and Realistic Image Synthesis 208

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING

Chapter 8

Meshing
The general strategies for automatic meshing described in the previous chapters
determine where and when to subdivide, but not how to subdivide. The actual
mechanics of subdividing geometry will be addressed in this chapter.
Basic subdivision techniques are addressed first. These are useful both for
producing an initial uniform mesh as well as for further refining a mesh by
adaptive subdivision. A priori algorithms for determining mesh boundaries that
correspond to discontinuities in the radiosity function will also be discussed.
Meshing algorithms are best constructed on an underlying topological data
structure that efficiently represents the adjacencies between nodes, edges and
elements. The basic characteristics of such data structures and how they can be
applied to meshing will be described.
Finally, several alternatives to meshing have been developed in order to
avoid some of the complexities of surface subdivision. These approaches will
be examined both for their own sake as well as for the somewhat different light
they shed on the overall problem of approximating the radiosity function.

8.1 Basic Subdivision Techniques

A wide variety of subdivision techniques has been developed for finite and
boundary element analysis. Surveys have organized these techniques according
to a variety of taxonomies [92, 146, 211]. For our purposes, subdivision tech-
niques are broadly classified according to whether or not the mesh topology is
predetermined.
Algorithms that use a predetermined topology subdivide by mapping a mesh
template to the geometry. The template is a predefined subdivision into standard
elements, such as a rectangular grid. The essential step of the meshing process
consists of determining a mapping between the geometry and the template. The
alternative, in which the topology is not predetermined, is to decompose the
geometry into elements by locating and connecting nodes according to some
procedure. As a trivial example, a polygon might be subdivided by connecting
its vertices to form triangles.

Radiosity and Realistic Image Synthesis 209

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.2 MESH TEMPLATE METHODS

Figure 8.1: Subdivision using a mesh template.

Both basic subdivision approaches have been used for radiosity and are
described further in the following sections.

8.2 Mesh Template Methods

8.2.1 Grid Superposition

The most direct way of mapping a mesh template to the geometry is simply to
superimpose the template on the geometry. The geometry is then split along
the grid lines to create the mesh (see Figure 8.1). The chief advantage of
this approach is that it generates perfectly regular, well-shaped elements on the
interior of the geometry.
Unfortunately, the cookie-cutter approach frequently generates unaccept-
able elements where the template intersects the boundary. Elements are often
poorly shaped, with low aspect ratios, concavities or extra vertices. Element
shape can be improved by smoothing the mesh (as in section 8.4), but smooth-
ing cannot eliminate extra vertices.
Grid superposition schemes improve on the cookie-cutter approach by pro-
viding special handling for elements near the boundary. For example, elements
of the template that contain pieces of the boundary may be partially collapsed
to eliminate extra vertices.
Baum et al. [18] describe an alternative in which the mesh template is used
to create only those elements that lie fully inside the boundary. The region
between the boundary and the interior elements is then triangulated. 1
1
This subdivision technique was first described by Rockwood as a means of tesselating
trimmed spline surfaces for display [194].

Radiosity and Realistic Image Synthesis 210

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.2 MESH TEMPLATE METHODS

Figure 8.2: Subdivision using a template for interior elements and triangulation
to connect to boundary.

In the first step of Baums approach the template is superimposed on the

geometry. The grid is traversed and elements are created only where all four
corners of the grid square are well inside the boundary. Grid corners that are
inside but very close to the boundary are rejected, since they may cause poorly
shaped elements during the triangulation step. Next, the boundary is subdivided
at regular intervals corresponding to the desired mesh density. Finally, the region
between the subdivided boundary and the internal elements is triangulated (see
Figure 8.2).
Since the triangulation will link element edges to all boundary vertices, small
boundary features are incorporated into standard three-sided elements. However,
if the boundary contains features or edges that are much smaller than the target
element size, triangulation will often result in poorly shaped elements and a
poorly graded mesh (see Figure 8.3).
In general, the graceful incorporation of small boundary features requires a
nonuniform mesh. Producing a well-graded, nonuniform mesh requires greater
control over element size than is provided by a uniform template. However,
template methods can be generalized to produce nonuniform meshes by using
an adaptive, variable-size template [211].

8.2.2 Template Mapping

The use of a template is simplified if the template can be mapped to fit the
geometry exactly. For example, a rectangular template can be mapped exactly

Radiosity and Realistic Image Synthesis 211

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.2 MESH TEMPLATE METHODS

Figure 8.3: Small features of the geometry may cause poorly shaped elements.

to fill a convex quadrilateral. The mapping in this case is the same as for the
bilinear isoparametric element described in Chapter 5. Once the mapping is
obtained, a set of elements corresponding to the template can be turned into an
actual mesh simply by transforming the parametric coordinates for each node to
geometric coordinates. More general mappings can be used to handle regions
with curved boundaries or with more than four sides [235], although these have
not been applied to radiosity to date.
Although mapping increases the range of geometries that can be meshed
with template subdivision, there are limitations. The mapping may introduce
severe distortions, resulting in a nonuniform mesh of poorly shaped elements.
In addition, although a conformal mapping does exist between any two simply
connected regions with an equal number of vertices, directly mapping geometries
with more than four sides can be complicated.

8.2.3 Multiblocking
An alternative when the geometry is too complicated or is concave is to subdivide
the geometry into simpler regions or blocks first. Templates can then be mapped
directly to each of the regions.
Cohen et al. [61] use multiblocking with mapped templates in an early
radiosity implementation. Surfaces are subdivided into convex quadrilaterals
and triangles by connecting vertices of the geometry These regions are then
subdivided into quadrilateral or triangular elements by splitting the edges in half

Radiosity and Realistic Image Synthesis 212

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.2 MESH TEMPLATE METHODS

Figure 8.4: Subdivision with mapped elements. The geometry is first split into
simple shapes. In Cohens algorithm, the boundaries are then bisected and new
vertices joined to form elements. Subdivision proceeds recursively as needed.

and joining them across the interior to create four elements. Finer mesh densities
are obtained by continuing the subdivision recursively until the desired element
size is reached (see Figure 8.4). Splitting edges at the midpoint guarantees that
new elements created on either side of the edge will be conforming.
When the initial blocks are obtained by joining the vertices of the original
geometry, the blocks may sometimes have extreme shapes that produce highly
distorted mappings. These can be avoided by adding new vertices on the bound-
ary or the interior of the geometry. As noted in [146], however, once such an
algorithm is available for the initial subdivision, there is little reason not to use
it to produce the entire mesh.

Radiosity and Realistic Image Synthesis 213

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.2 MESH TEMPLATE METHODS

8.2.4 Adaptive Subdivision with Templates

Once the mapping has been established, a mesh produced using a template can be
refined during adaptive subdivision by applying the template again to elements
requiring further subdivision. In Cohens quadtree subdivision algorithm, for
example, quadtree subdivision is invoked again following the initial solution to
refine elements where needed.
As a means of a posteriori mesh refinement, mapped templates (of which
the quadtree scheme is the most common example for radiosity) have the dis-
advantage that the placement of element boundaries is inflexible. New element
edges cannot be oriented to distribute the error as evenly as possible. As a result,
adaptive subdivision using template mapping may require more elements than
necessary to achieve a certain accuracy.
Template mapping approaches, like the quadtree algorithm, also tend to
generate nonconforming elements in regions of changing mesh density. For ex-
ample, when neighboring elements are subdivided to different levels, quadtree
subdivision creates T-vertices. If the node at the T-vertex is not shared by the
larger neighboring element, a discontinuity in the approximation will be intro-
duced along that edge, since the interpolated value at that point will generally
be different from the value computed at the node for the same point (see Fig-
ure 8.5). On the other hand, if the node at the T-vertex is shared with the larger
element, the large element is no longer well shaped, since it contains 5 nodes,
one of which is at the apex of edges forming a l80-degree angle.
Quadtree algorithms have employed several techniques for handling T-
vertices Cohen et al. create special nodes at T-vertices, sometimes called slave
nodes in the finite element literature. The function is not actually evaluated at
slave nodes. Instead, the value at a slave node is interpolated (linearly, in this
case) from the nodes of the largest element sharing that edge (nodes A and B in
Figure 8.6). This allows interpolation on the larger element to ignore the slave
nodes without introducing discontinuities.
Baum, et al. [18] describe an alternative way of handling T-vertices. Nodes
at T-vertices are shared by all elements surrounding the vertex and the function
is evaluated at the node. To maintain correct element shape, the larger element
is then triangulated to form smaller properly shaped elements (see Figure 8.7).
The resulting modified hierarchy is called a tri-quadtree.
The extra triangulation in the tri-quadtree is not a permanent part of the
mesh. The triangulation of a problem element is used only when needed for lin-
ear interpolation. If further adaptive subdivision is required later in the solution,
it will be performed on the original element, not the triangulation.
Adaptive subdivision using a quadtree often produces poorly graded meshes.
Figure 8.8 shows an example in which a large element is adjacent to a highly

Radiosity and Realistic Image Synthesis 214

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.2 MESH TEMPLATE METHODS

Figure 8.5: Discontinuity introduced by incorrectly handled T-vertices

subdivided region. Triangulation of the large element in this case will produce
triangles with a poor aspect ratio. The use of slave nodes may produce a visible
artifact along the long edge, due to its close proximity to nodes at which the
function is actually evaluated.

Better mesh grading can be achieved by requiring a balanced or restricted

quadtree [18, 211]. Neighboring elements in a restricted quadtree are allowed
to differ by only one level of subdivision. A balanced tree can be achieved by
checking the neighbors of an element when it is subdivided, and by subdividing
those neighbors recursively.2

2
Algorithms for performing this and other operations on quadtrees can be found in
Samet [204].

Radiosity and Realistic Image Synthesis 215

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.3 DECOMPOSITION METHODS

Figure 8.6: T-vertices treated as slave nodes.

Figure 8.7: Tri-quadtree used to eliminate t-vertices.

8.3 Decomposition Methods

Mapped template methods are efficient and produce well shaped elements for
simple geometries. However, their limited flexibility is a drawback when sub-
dividing more complicated geometries and during a posteriori mesh refinement.
Subdivision methods that decompose the geometry into elements piece by
piece provide greater flexibility. There are two basic approaches to this decom-
position: the nodes and elements can be produced at the same time using a
single procedure or the nodes can be produced first and then connected to form
edges in an independent step.

Radiosity and Realistic Image Synthesis 216

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.3 DECOMPOSITION METHODS

Unrestricted quad-tree Restricted quad-tree

Figure 8.8: Use of restricted quadtree to improve mesh grading.

8.3.1 NodesElementsTogether Decomposition

In this approach, the geometry is subdivided by creating nodes and joining
them with edges to generate new elements one by one. Subdivision can be
performed recursively, first subdividing the entire geometry into initial elements,
then splitting those elements again until the desired element size is reached.
During subdivision, new nodes can be positioned according to criterion selected
to produce a mesh with certain density or smoothness properties. Chew [53]
and Watson [256] describe incremental Delaunay triangulation algorithms of
this type. Alternatively, subdivision may proceed from the boundary inward,
splitting off triangles one by one until the entire geometry is meshed.

8.3.2 Decomposition by Recursive Splitting

Campbell and Fussell [43] subdivide surfaces by recursive splitting using a two-
dimensional binary space partition (BSP) tree. The BSP-tree allows arbitrary
orientation of element edges (see Figure 8.9). Elements are created by splitting
larger regions along an arbitrary edge. The ability to incorporate arbitrarily
oriented edges allows Campbell and Fussel to include previously determined
discontinuity boundaries into the mesh.
Campbell and Fussell also use the flexibility of BSP subdivision advan-
tageously during adaptive subdivision. The edge used to split an element is
oriented to minimize the variation in radiosity over the new elements. Using
optimization techniques, Campbell and Fussell locate the global maximum and
minimum for the element and subdivide halfway between the two extreme points

Radiosity and Realistic Image Synthesis 217

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.3 DECOMPOSITION METHODS

Figure 8.9: Hierarchical surface subdivision using a BSP tree (after Campbell,
1990).

along an edge perpendicular to the line joining the extrema.

Although the BSP-tree is more flexible than the quadtree, it is also more
difficult to maintain conforming elements with the bsp-tree. T-vertices occur
frequently, since neighboring elements are split along edges that are not likely to
meet at a common node. Campbell and Fussell treat T-vertices as slave nodes.
They also recommend a final clean-up pass following adaptive subdivision, dur-
ing which the tree is balanced by subdividing elements with highly subdivided
neighbors and those with a poor aspect ratio.
Lischinski and Tampieri [154], who also use a bsp-tree to represent surface
subdivision, avoid the clean-up pass by simultaneously maintaining a data struc-
ture representing the mesh topology. Nodes of the BSP-tree point to edges in
a winged-edge data structure. T-vertices can then be eliminated using a simple
trianlgulation pass.

8.3.3 Decomposition by Advancing Front

In contrast to recursive splitting, decomposition may also proceed from the
boundary inward, splitting off finished elements one at a time. This is some-
times called the advancing front method, because the forward boundary of the
meshed region is progressively advanced into the unmeshed region until the en-
tire geometry is subdivided [92]. Algorithms of this type are distinguished by
the technique with which the front is advanced. For example, the front may
be searched to locate the smallest angle, which is then incorporated into a new
triangular element and split off from the unmeshed region (see Figure 8.10). The
advancing front method is also known as paving when it is used to subdivide

Radiosity and Realistic Image Synthesis 218

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.3 DECOMPOSITION METHODS

Figure 8.10: A intermediate stage of an advancing front algorithm.

into quadrilateral elements.

The advancing front method allows control over element quality, since qual-
ity criteria can be explicitly incorporated into the rules according to which ele-
ments are split off from the front. However, this approach can run into difficul-
ties for complicated geometries. For example, backtracking may be required to
undo decisions, resulting in fronts for which no good candidates are available
for splitting. The advancing front technique has not yet been applied to radiosity
meshing.

8.3.4 Nodes-First Decomposition

In a nodes-first method, decomposition is accomplished by first positioning the
nodes and then connecting them to form edges. Because nodes are placed and
connected in independent steps, the procedures for each task can be chosen
independently from a greater range of possibilities. This leads to greater overall
flexibility. Nodes can be placed anywhere, with nonuniform density if desired,
either to handle small features of the geometry, or to address some anticipated
behavior of the function. All the nodes may be laid out before any edges are
created or the subdivision may proceed recursively by creating some nodes and
linking them, then creating more nodes within the resulting elements, and so on.
Figure 8.11 shows an example in which nodes are generated by superimposing
a grid over the geometry and placing a node more or less randomly in each grid
cell.
The nodes are usually connected by triangulation. Delaunay triangulation 3

Radiosity and Realistic Image Synthesis 219

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.3 DECOMPOSITION METHODS

Figure 8.11: Subdivision using a nodes-first algorithm.

is commonly used since it guarantees triangles that are as well shaped as possible
(according to a certain criterion) for a given set of nodes. Delaunay triangulations
are often also constructed incrementally, positioning new nodes in relationship
to existing elements in order to achieve elements with desired shape or density.
Algorithms for connecting nodes into quadrilaterals are less common.
For radiosity applications, Heckbert [121] has incorporated a triangulation
approach into an a priori discontinuity meshing scheme. Critical element bound-
3
A Delaunay triangulation connects a given set of points so that the circle circum-
scribing any triangle contains only the points belonging to that triangle (except in the
case where four or more points are cocircular). A Delaunay triangulation also minimizes
the smallest angle of its triangles over all possible triangulations of the set of points and
thus avoids thin slivers as far as possible. A further, not inconsequential, advantage of
Delaunay triangulations is that they are a generally useful construct for computational
geometry and algorithms for producing them have thus received great attention. The
standard computational geometry text by Preparata and Shamos [184] provides a detailed
discussion. The article by Schumaker [207] provides a short overview and a discussion
of practical issues.

Radiosity and Realistic Image Synthesis 220

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.4 MESH SMOOTHING

aries corresponding to discontinuities in the radiosity function are placed first.

The remainder of the mesh is determined by placing and connecting nodes us-
ing a constrained Delaunay triangulation, which preserves existing edges [53]
(Heckbert uses an algorithm based on [33]). The flexibility of decomposition
by triangulation is particularly useful when dealing with the complex region
boundaries often created by discontinuity meshing. Lischinski et al. [154] have
also used triangulation to incorporate discontinuity edges.
Sturzlinger [227] employs a variation on the nodes-first approach designed
specifically for a radiosity implementation in which constant elements are used
during the solution and linear interpolation is used during rendering. The initial
mesh is created by first positioning nodes according to a uniform grid. Constant
elements consisting of the Voronoi polygonalization of the nodes are used during
the solution. Prior to rendering, the same nodes are triangulated to provide
elements suitable for linear interpolation. The Voronoi diagram is the straight-
line dual of the Delaunay triangulation, which makes conversion between the
Voronoi mesh and the triangular mesh easy and convenient. Sturzlinger also uses
incremental refinement of the Voronoi polygonalization to implement adaptive
subdivision.

8.4 Mesh Smoothing

The quality of the subdivision produced by both template and decomposition
methods can often be improved by smoothing the mesh. Mesh smoothing con-
sists of several passes of relaxation during which nodes are repositioned to
improve element shape and grading. In each relaxation pass each node is typi-
cally relocated to the centroid, P, of the n adjacent nodes located at Pi using a
formula like the following:
n

P
1
P = i (8.1)
n i=1

Not all nodes are equally free to move. Nodes along fixed boundaries can
move only along the boundary. Fixed boundaries include the original edges of
the geometry and mesh boundaries that may have been added a priori, such as
discontinuity boundaries. Nodes at the intersection of fixed boundaries are not
free to move at all.
Mesh smoothing is a useful tool because it relieves the subdivision algo-
rithms of some responsibility for element shape. The subdivision algorithm is
then free to concentrate on creating a correct topology. For radiosity, mesh
smoothing has been employed by Heckbert [121] as a final step following tri-
angulation.

Radiosity and Realistic Image Synthesis 221

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Mesh relaxation can also be used for a posteriori mesh refinement (the r-
refinement method described in Chapter 6) by incorporating local element error
into the relocation method. For example, the following relocation formula moves
the node P according to the area weighted errors of the n elements adjacent to
the node:
i=1 xi ei Ai
n
P = (8.2)
i=1 Ai
n

where n is the number of elements adjacent to P, xi is the centroid of adjacent

element i, ei is the approximation error for element i, and Ai is the area of
element i. This formula will tend to move the node in a direction that equalizes
the error among the elements adjacent to the node.

8.5 Discontinuity Meshing

The radiosity function is piecewise smooth (C) within regions bounded by
discontinuities of various orders. The failure to resolve these discontinuities
correctly can result in highly visible artifacts, as demonstrated in Chapter 6.
A posteriori refinement is not very effective at reducing error in the neighbor-
hood of discontinuities. Since the basis functions are continuous, discontinuitis
can only be introduced into the approximation along element boundaries. Re-
ducing error in the neighborhood of a discontinuity requires either a relatively
high mesh density, which is expensive, or exact alignment of element edges with
the discontinuity, which is difficult to achieve using an a posteriori algorithm.
However, it is possible to determine the location of discontinuity boundaries
a priori. Discontinuities in the radiosity function correspond to transitions in
occlusion between source and receiving surfaces. These are purely geometric in
nature and can be determined before radiosities are computed.

8.5.1 Discontinuities in Value

A discontinuity in the value of the radiosity function occurs; when one surface
touches another (see Figure 8.12).4 The discontinuity is caused by the abrupt
transition from complete visibility to complete occlusion that is encountered in
crossing the line of contact between the surfaces. The locations of these value
discontinuities can be identified by determining the geometric intersections of all
surfaces, as described by Baum, et al. [18]. Although conceptually straightfor-
ward, the intersection computation requires careful handling to avoid numerical
difficulties. Isect, the program used by Baum to resolve surface intersections,
4
Value discontinuities also occur at the boundaries of shadows cast by point light
sources. Point lights are discussed in the context of the radiosity method in Chapter 10.

Radiosity and Realistic Image Synthesis 222

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.12: A discontinuity in value. The plot to the lower right shows the ra-
diosity of the surface as a function of position along the horizontal line overlayed
on the upper right image. Courtesy of Daniel Lichinski and Filippo Tampieri,
Program of Computer Graphics, Cornell University.

was designed to compute the intersection of solid model boundary representa-

tions robustly [209].
Once boundaries have been identified, they can be inserted into the repre-
sentation of the polygon as specially flagged edges. The insertion of new edges
is easier if polygons are represented by a robust topological data structure, like
the winged-edge data structure (topological data structures are discussed in sec-
tion 8.6). When the polygon is meshed, the discontinuity edges become element
boundaries. Correct reconstruction of the discontinuity requires that adjacent
elements do not share nodes along such boundaries.
In a point collocation radiosity method, in which form factors are evaluated
at the nodes, occlusion testing must be handled carefully along value disconti-
nuities. As shown in Figure 8.13, the geometric location of a node on such a
boundary is not sufficient in itself to determine its occlusion with respect to the
intersecting surface. Such nodes actually represent the behavior of the radiosity
function in the limit moving towards the boundary from the left or the right.
Any purely numerical shadow algorithm, such as the Z-buffer or ray tracing,
will return the same answer for nodes with the same location (in most cases it
will ignore occlusion by the intersecting surface).
One solution is to move the node temporarily away from the boundary by
a small amount in the appropriate direction. It is also possible to determine the
occlusion with respect to the intersecting surface topologically. If the source
and the element to which the node belongs are on opposite sides of the plane of
the intersecting surface, the node is in shadow. If they are on the same side, the

Radiosity and Realistic Image Synthesis 223

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.13: Occlusion testing at a discontinuity in value. a) A numerical shadow

test will return the same values for nodes a and b, since they have the same geomet-
ric location. b) Occlusion with respect to the intersecting surface is determined
by whether the corresponding element lies on the same side or the opposite side
of the plane from the light source.

node is in the light (or, at least, not shadowed by the intersecting surface). When
the light straddles the plane of the occluder, the test remains straightforward for
point sampling from factor algorithms (e.g., ray traced form factors), since it is
performed for each point sample. For other algorithms, the light in the straddling
case may have to be split temporarily across the plane.
The image in Figure 8.14 shows artifacts typically produced when value
discontinuities are ignored. Not also the extra elements generated by adaptive
subdivision in an unsuccessful attempt to resolve the discontinuity. Compare
this image with Figure 8.15, in which discontinuity boundaries are incorporated
into the mesh. The shadow leak on the wall behind the table top is eliminated
and no adaptive subdivision is required along a boundary.

8.5.2 First and Second Derivative Discontinuities

Discontinuities in the first and second derivatives of the radiosity function result
from qualitative transitions in occlusion between surfaces in a polygonal environ-
ment.5 The image in Figure 8.16 shows the geometry of occlusion for a simple
scene. The resulting first and second derivative discontinuities are evident in the
5
Additional discontinuities of arbitrary degree will occur due to interreflection, as
Heckbert demonstrates [120], but these are of much lesser visual consequence.

Radiosity and Realistic Image Synthesis 224

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.14: Artifacts due to ignoring a discontinuity in value.

Figure 8.15: Discontinuity in value correctly handled.

Radiosity and Realistic Image Synthesis 225

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.16: First and second derivative discontinuity boundaries for a polyg-
onal environment. Note the visible discontinuity in the first derivative of the
radiosity function plotted in the lower right of the figure. Courtesy of Daniel
Lischinski and Filippo Tampieri, Program of Computer Graphics, Cornell Uni-
versity.

plot of the radiosity function. As apparent in this image, discontinuities define

the outer boundary of the shadow penumbra and the boundary of the umbra.
Additional discontinuities occur within the penumbra itself.
In a polygonal environment, certain geometric events cause a qualitative
transition in the occlusion of an illumination source. The variation in the occlu-
sion of a source undergoes a qualitative transition at certain geometric events.
Imagine viewing a partially occluded source from a point moving along a re-
ceiving surface. As the viewpoint moves, more of the source is revealed (see
Figure 8.17). In general, the visibility of the source (and thus the direct energy
transfer for a constant source) varies quadratically according to a function de-
termined by the relative orientation of the overlapping edges of the source and
the occluding polygon.
Whenever an edge joins or leaves the set of overlapping edges, a new
quadratic variation is determined and there is a discontinuity in the first or second
derivative. This transition is called a visual event. There are two classes of visual
events, Vertex-Edge (VE or EV) and Edge-Edge-Edge (EEE), corresponding to
the two possible ways that an edge can leave or join the set of overlapping
edges.
A VE event occurs when a vertex of the source crosses an edge of the
occluding polygon (first row of Figure 8.17), or conversely when an edge of
the source crosses a vertex of the occluder (for convenience this case may be
differentiated as an EV event). The event defines a critical surface, that is, the

Radiosity and Realistic Image Synthesis 226

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.17: VE and EEE visual events, from a viewpoint looking back toward
the occluded light source.

wedge formed by the vertex and the edge that cause the event. The intersection
of this wedge with the scene polygons determines the discontinuity boundaries
resulting from this event. The critical surface and its effect on the radiosity
function are shown in Figures 8.18 and 8.19.
A discontinuity in the first derivative occurs when an edge of the source and
an edge of the occluder are coplanar, as shown in the second row of Figure 8.17.
First derivative discontinuities are evident in Figure 8.16.
Visual events can also occur where neither the vertex nor the edge belongs
to the source, as shown in row three of Figure 8.17. The visual event involving
the two occluding polygons in this case modifies the function describing the
change in source visibility, even though the event does not involve any of the
source vertices or edges.
The other class of visual events, EEE events, involves transitions caused by

Radiosity and Realistic Image Synthesis 227

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.18: A VE event caused by a source vertex and an occluder edge. Cour-
tesy of Daniel Lischinski and Filippo Tampieri, Program of Computer Graphics
Cornell University.

Figure 8.19: A VE event caused by an occluder vertex and a source edge.

Lischinski et al. differentiate these as EV events. Courtesy of Daniel Lischinski
and Filippo Tampieri, Program of Computer Graphics, Cornell University.

overlapping edges of multiple occluders (fourth row of Figure 8.17). The critical
surfaces in this case are quadric and the resulting discontinuity boundaries are
curves.
The computational challenge in discontinuity meshing is to test the scene
polygons against the critical surfaces efficiently and robustly and insert the re-
sulting boundaries into the polygon mesh.

Radiosity and Realistic Image Synthesis 228

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.20: Determining penumbra and umbra volumes using Nishita and
Nakamaes algorithm. (After Nishista and Nakamae, 1985).

8.5.3 Shadow Volume Algorithms

Early algorithms deal only with the subset of events that define the boundaries of
the umbra and penumbra. Nishita and Nakamae [175] determine these bound-
aries by computing shadow volumes formed by an object and each vertex of
the light polygon (see Figure 8.20.) The umbra volume for a single occluder
is defined by the intersection of the shadow volumes. The penumbra volume is
defined by the three-dimensional convex hull containing the shadow volumes.
The intersection of the umbra and penumbra volumes with surfaces defines the
penumbra and umbra boundaries. This approach ignores discontinuities within
the penumbra.
Campbell [42] also resolves only the outer penumbra and umbra bound-
aries. Like Nishita and Nakamae, Campbell constructs umbra and penumbra
volumes for each source-occluder pair. However, he avoids computing a three-
dimensional convex hull by constructing the volumes directly from critical sur-
faces.
Campbells algorithm assumes convex source and occluder polygons. For
each edge of the occluder, or blocker, there exists a critical surface formed with
respect to each vertex of the source. The minimum blocker extremal plane is
defined to be the critical surface that forms the minimum angle with respect to the
plane of the blocker (see Figure 8.21). A minimum blocker extremal plane exists
for each occluder edge. Likewise, there is a minimum source extremal plane
for each edge of the source. The penumbra volume consists of the intersection

Radiosity and Realistic Image Synthesis 229

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.21: The minimum and maximum extremal planes and the penumbra
volume created using the minimum blocker and source extremal planes. (After
Campbell, 1991).

of the negative halfspaces of all of the minimum blocker and source extremal
planes (see Figure 8.20). The umbra volume is constructed similarly, using the
maximum blocker extremal planes.

The resulting umbra and penumbra volumes are stored as BSP-trees. These
volumes are then merged with a BSP-tree representing the unified penumbra
and umbra volumes for the entire scene to provide an efficient means of testing
polygons against the volumes. Chin and Feiner [54] first describe the use of a
BSP-tree to represent merged shadow volumes, but for point lights only. The
merging of occlusion and umbra volumes for area lights requires more general
methods, for which Campbell uses algorithms developed by Naylor [171]. Later
work by Chin [55] describes a similar generalization of the BSP-tree shadow
volume technique to area lights.

After constructing the shadow volumes, Campbells algorithm tests each

polygon in the scene against the shadow BSP-tree. Polygons are split where they
cross planes defining shadow volumes, thus effectively inserting the discontinuity
edges. Campbells approach allows all regions to be classified as totally in
shadow, totally in light, or partially occluded. Shadow testing can be eliminated
for the subsequent computation of form factors for nodes within regions classified
as totally in the light or totally in shadow. Figure 8.22 shows an example of a
mesh produced by Campbells shadow algorithm.

Radiosity and Realistic Image Synthesis 230

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.22: A mesh produced by Campbells penumbra volume algorithm.

Courtesy of A. T. Campbell III, University of Texas.

8.5.4 Critical Surface Algorithms

Heckbert [121] and Lischinski et al. [154] have independently developed similar
algorithms that compute discontinuity boundaries due to VE (and EV) events
involving the vertices or edges of source polygons. For a given source polygon,
the wedge for each VE event is formed and intersected with the scene polygons.
These intersections determine discontinuity edges, which are then inserted into
the mesh data structure for the polygon.
For the intersection step, Heckbert tests the wedge against every polygon
in the scene (see Figure 8.23). Each intersection forms a two-dimensional span
on the wedge surface. A two-dimensional hidden span algorithm determines the
visibility of the spans with respect to the wedge vertex. The visible portions
of the spans correspond to discontinuity boundaries. These are accumulated
for each polygon. Following the intersection step, a line-sweep algorithm is
performed on each polygon to connect the discontinuity edges at coincident
endpoints, split overlapping edges where necessary, and assemble the result
into a connected winged-edge representation. The steps of this algorithm are
diagramed in Figure 8.23. Heckbert performs discontinuity meshing for emitters
only, on the assumption that discontinuities due to reflected light contribute very
little to error in the approximation.
Lischinski, et al. store the scene polygons as a BSP-tree. The construction
of the BSP-tree requires additional processing and storage, but has the advantage
of allowing the polygons to be visited in front-to-back order when testing for
intersections against the wedge. Each intersection of the wedge with a polygon
generates a discontinuity edge and clips that portion of the wedge. When the
wedge is entirely clipped away, no further polygons need be tested.
Each polygons mesh is stored as a two-dimensional BSP-tree. As dis-
continuity edges are generated, they are filtered down the BSP-tree until they
reach leaf nodes, at which point the corresponding faces of the meshed polygon
are split. Lischinski also maintains a winged-edge representation of the mesh

Radiosity and Realistic Image Synthesis 231

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.23: Steps in Heckberts computation of discontinuity edges.

Figure 8.24: Steps in the method of Lischinski et al. for the computation of
discontinuity edges.

Radiosity and Realistic Image Synthesis 232

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.5 DISCONTINUITY MESHING

Figure 8.25: Artifacts resulting from conventional quadtree-based adaptive sub-

division. Courtesy of Daniel Lischinski and Filippo Tampieri, Program of Com-
puter Graphics, Cornell University.

Figure 8.26: Discontinuity Meshing. Courtesy of Daniel Lischinski and Filippo

Tampieri, Program of Computer Graphics, Cornell University.

topology, with each node of the tree pointing to an edge in the winged-edge
data structure. The steps of Lischinskis algorithm are shown in Figure 8.24.
Figures 8.25 and 8.26 compare results for conventional quadtree meshing and
discontinuity meshing using this algorithm. Color plates 32 and 33 show images
computed with and without Lischinskis algorithm.
Heckbert and Lischinski ignore EEE events and VE events involving mul-
tiple occluders. Handling EEE events correctly is an involved problem. Al-
gorithms applicable to finding discontinuity boundaries due to EEE events are

Radiosity and Realistic Image Synthesis 233

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.6 TOPOLOGICAL DATA STRUCTURES AND OPERATORS

Figure 8.27: Representation of mesh topology using an embedded graph.

described by Teller [233] for the case of a sequence of convex holes.

8.6 Topological Data Structures and Operators

The representation of a mesh naturally includes data describing node locations,
surface normals, material attributes, and so on. In addition to geometric data,
a complete representation also includes the mesh topology, which describes the
adjacency between mesh elements, element edges, and nodes.
Many of the meshing algorithms described in this chapter depend on knowl-
edge of adjacency. For example, splitting one element during adaptive subdivi-
sion often requires inserting a new node on an edge shared by adjacent elements.
Computing a gradient at a node may require visiting adjacent nodes. A data
structure that maintains the adjacency, or topology, of these entities explicitly
can greatly improve the efficiency and simplicity of the algorithms described in
this chapter.
The topology of a mesh can be represented by an embedded graph, that is,
a graph that has been mapped to a surface. The faces of an embedded graph
are the regions of the surface bounded by the graph edges. The graph vertices
correspond to mesh nodes, edges correspond to element boundaries and faces
correspond to elements (see Figure 8.27). When the meshed surface does not
form the closed boundary of a solid, there are no elements adjacent to the outer
edges of the mesh and a pseudo-face is often assumed in order to maintain
topological consistency.

Radiosity and Realistic Image Synthesis 234

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.6 TOPOLOGICAL DATA STRUCTURES AND OPERATORS

8.6.1 Data Structure Criteria

A variety of data structures for representing the adjacency graph has been de-
veloped for solid modeling. Several are described and evaluated by Weiler in
[261, 262] and by Mantyla in [158]. The data structures compared by Weiler
are edge-based; all information required to find adjacent vertices, edges or faces
is contained in the edge record. The winged-edge data structure, first developed
by Baumgart for computer vision [22], may be the most familiar example, but
many others have also been devised.
Topological data structures can be compared on the basis of storage and
performance. Evaluation of the storage required to represent a given number of
edges (or vertices) is straightforward. (A comparison of the storage requirements
for several data structures, including the winged-edge, can be found in [6, 261].)
However, compactness must be weighed against performance; an extremely
succinct data structure may require extra pointer manipulations to extract certain
information.
Performance is less straightforward to evaluate. The number of field ac-
cesses required for various operations is often used as a measure. Performance
also depends on how often operations access different records, since these may
reside on different pages in virtual memory. Papers by Ala and Woo contain
experiments and discussions that help clarify these issues [6, 268]. However,
actual performance will depend strongly on the mix of operations typical of
the application. A careful evaluation of data structures would require statistical
measurements of the operations performed by the particular application.
To date no such evaluation of topological data structures has been made
for radiosity meshing. One reason is undoubtedly that the cost of numerical
computation (integration, interpolation, etc.) tends to overwhelm other costs.
Storage, on the other hand, is of immediate practical concern. Image synthesis
is often a small component of a larger application, such as an architectural or
industrial design package, and the additional memory requirements of radiosity
are a serious consideration.

8.6.2 The Winged-Edge Data Structure

The winged-edge data structure (WED) will be used to illustrate how a topo-
logical data structure can be used to represent a mesh. The winged-edge data
structure is fairly compact, provides reasonable performance, and has often been
used for radiosity [61, 121, 154].
Figure 8.28 contains a diagram of the basic winged-edge data structure. As
the name implies, the basic record is that of an edge. The record points explicitly
to the pair of faces that are adjacent to the edge. It also points to the two vertices
that are the endpoints of the edge (in order, thus orienting the edge). At each

Radiosity and Realistic Image Synthesis 235

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.6 TOPOLOGICAL DATA STRUCTURES AND OPERATORS

typedef struct {
stmct edge *vertex_edge ;
} VERTEX ;

typedef struct {
struct edge *face_edge ;
} FACE ;

typedef struct edge {

VERTEX *head_vertex ;
VERTEX *tail_vertex ;
FACE *right_face ;
FACE *left_face ;
struct edge *cw_right_edge
struct edge *ccw_right_edge ;
struct edge *cw_left_edge ;
struct edge *ccw_left_edge ;
} EDGE ;

Figure 8.28: The winged-edge data structure.

endpoint, the next edge (or wing) around each of the adjacent faces is stored in
both the clockwise and counterclockwise directions. The data structures for a
face and a vertex need only keep a single pointer to any of the adjacent edges.
Operations on the topology represented by the winged-edge data structure
consist of either modifications to the topology or queries about adjacency re-
lationships. Modifications to the topology can be expressed in terms of Euler
operators [22]. (Good introductions are provided by [117, 158, 159, 267].) Euler
operators are so called because they maintain Eulers formula, which expresses
the relationship between the number of vertices, edges and faces in a topolog-
ically consistent adjacency graph. For a simple polyhedron with no holes this
formula is
VE+F=2 (8.3)

where V, E, and F are the number of vertices, edges, and faces,respectively. 6

A complete set of Euler operators provides for all possible modifications
to the topology, and guarantees that the topology following any operation will
be consistent. In addition, the use of operators simplifies the coding of higher
6
The more general formula is given by V E + F = 2 2g, where g is the genus
of the polyhedron. Generalizations also exist for other dimensions.

Radiosity and Realistic Image Synthesis 236

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.6 TOPOLOGICAL DATA STRUCTURES AND OPERATORS

Figure 8.29: Operations on the winged-edge data structure.

level meshing operations by encapsulating the manipulation of the underlying

data structures.
A complete set of Euler operators requires only five operators and their in-
verses [159]. Assuming no holes for the purpose of this discussion, these can be
reduced to three. The three operators, plus one more included for convenience,
are listed here:

1. Make-Vertex-Face (MVF) creates a new graph by creating a single vertex

and face.

Radiosity and Realistic Image Synthesis 237

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.6 TOPOLOGICAL DATA STRUCTURES AND OPERATORS

2. Make-Edge-Vertex (MEV) adds a vertex and connects it to an existing

vertex using a new edge.

3. Make-Edge-Face (MEF) connects two existing vertices to create a new

edge, which splits an existing face into two faces.

4. Split-Edge-Make-Vertex (SEMV) creates a new vertex and a new edge by

splitting an existing edge into two edges.7

Figure 8.29 shows how a simple quadtree meshing scheme could be imple-
mented using Euler operators. A graph representing the original geometry is
first created using an initial MVF operation, followed by several calls to MEV,
and finishing with MEF to close the polygon. (One of the two resulting faces is
a pseudo-face that represents the exterior region of the graph). The polygon
is then split along the vertical and horizontal directions using SEMV to split
edges and MEF to split faces by linking vertices with new edges. Finally, one
element of the initial mesh is split further, showing how the use of a topolog-
ically complete data structure correctly handles the insertion of the new vertex
(labeled A in the diagram) along the shared edge.
Once the mesh has been created, higher level algorithms will need to query
the topological data structure for adjacency information. For example, various
ways of traversing the topology can be thought of as queries: get next face,
get next vertex around face, get next edge around vertex, and so on. The
implementations are straightforward in most cases, although edge orientation
requires special handling. When traversing the boundary of a face in a winged-
edge data structure for example, edges may be oriented arbitrarily. Therefore,
comparisons are required to determine which endpoint of the edge is the next
vertex around the face.
The basic topological data structure can be augmented with extra information
to speed up certain queries. Back-pointers might be added to allow direct access
to the parent object record from the vertex record. An extra pointer can be
added to the vertex record, allowing nodes to be linked into a list that can be
traversed directly during the solution, when it is necessary to visit all the nodes
of the scene in sequence to compute form factors or to update radiosities.
Finally, certain algorithms require the imposition of a mesh hierarchy. Hi-
erarchical data structures, like quadtrees or BSP-trees, can be maintained in
parallel with the topological data structure, with nodes of the tree pointing to
faces of the topology.

7
A helpful approach to implementing winged-edge operations is described in [98].

Radiosity and Realistic Image Synthesis 238

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.7 ALTERNATIVES TO MESHING

Figure 8.30: Approximation of the radiosity function using the rex data structure.
Each square on the shaded floor corresponds to a cell in the rex. Courtesy of
Paul Heckbert, University of California, Berkeley.

8.7 Alternatives to Meshing

The representation of the radiosity function by a mesh of elements requires the
subdivision of the model geometry, which is clearly a complicated process in
general. Several investigators have proposed that the representation of shading
be decoupled from the geometry as completely as possible to simplify algorithms
and data structures.
If nodes are uniformly distributed across the surface, they can be organized
as a simple array of sample points. An array of discrete samples representing
a continuous function suggests a texture map. Heckbert [119] has proposed
texture mapping as an alternative approach to storing and reconstructing an
approximation of the radiosity function [119].
In Heckberts approach the power arriving at a surface is sampled (using
Monte Carlo ray tracing) and stored at the pixels of a radiosity texture, or
rex. The connection between rex pixels and sample locations on the surface
is implicit in the mapping between texture coordinates and surface coordinates.
The rexes are organized into a quadtree to support adaptive sampling. During
rendering, the radiosity can be reconstructed from the rex samples using any
one of a wide selection of reconstruction kernels, just as in conventional texture

Radiosity and Realistic Image Synthesis 239

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.7 ALTERNATIVES TO MESHING

Figure 8.31: Use of the rex data structure with a higher sampling rate. Courtesy
of Paul Heckbert, University of California, Berkeley.

mapping. Heckbert uses linear interpolation and averaging to reconstruct the

radiosity function in image space.
An image computed using the rex algorithm is shown in Figure 8.30. In
Heckberts implementation, light and eye ray tracing are used to traces paths that
include specular reflection or transmission (such as for the lens in this image).
A low number of eye and light rays were used to compute the light transport in
this image. Statistical variance makes the structure of the rex visible. The image
in Figure 8.31 was computed using a higher number of light and eye rays.
The regularity of sampling imposed by rexes reduces the size and complexity
of the data structures since no explicit topology is maintained. Rather, adjacency
information needed during rendering is implicit in the quadtree.
The rex scheme limits the coupling of shading and geometry to the map-
ping between texture and surface coordinates. This is a particular advantage for
curved surfaces, which are turned into polygonal facets by most meshing algo-
rithms. However, determining an appropriate mapping is not always straight-
forward. Further, the regularity imposed by the rex does not provide the full
sampling flexibility provided by meshing. These limitations are ameliorated in
Heckberts approach by subdividing rexes until they are smaller than a certain
size in image space. The difficulties in sampling the shading function adequately
are then greatly reduced. For example, if the image resolution is known during

Radiosity and Realistic Image Synthesis 240

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 8. MESHING
8.7 ALTERNATIVES TO MESHING

sampling, discontinuities need not be determined exactly, since they will never
be resolved beyond pixel resolution during rendering.
Vedel [241] uses a sampling data structure similar to the rex. Vedels algo-
rithm takes advantage of the flexibility of reconstruction from an array samples
by computing the radiosity gradient and reconstructing using an elliptical filter
oriented perpendicularly to the gradient. The heuristic is that the sample val-
ues change most quickly in the direction of the gradient, and the reconstruction
kernel should thus give less weight to values in that direction.
A related example of the decoupling shading and geometry is provided by
Wards Radiance program [254]. In Wards algorithm, radiances computed us-
ing stochastic eye-ray tracing are cached in a spatial data structure (an octree).
Whenever shading is required at a new point, the octree is checked. If values are
stored within a close enough neighborhood, the new value is interpolated instead
of being computed directly. The spatial data structure is decoupled from the sur-
face geometry and mapping difficulties are avoided, since the world coordinates
of any surface point map directly into the spatial data structure without trans-
formation. Wards algorithm is specifically view-based, but a similar approach
might be applied to a view-independent radiosity approach.
Data structures that decouple shading and geometry may prove useful for
radiosity, particularly when image information is available to bound the required
sampling resolution. Published implementations for radiosity have been limited
to simple environments and thus many practical issues related to mapping and
sampling complicated geometries remain to be explored.

Radiosity and Realistic Image Synthesis 241

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING

Chapter 9

Rendering
Once a radiosity solution has been computed, the remaining step is to render
an image. During rendering, the model is projected onto the image plane, the
visible surfaces are determined, and a color is computed for each pixel. The
pixel color is derived from the radiance of the surface location visible at the
pixel. The radiance can be determined directly from the approximation of the
radiosity function computed during the solution. However, this approximation
is often not the best choice for rendering, due primarily to characteristics of the
human visual system. Previous chapters have addressed the computation of an
objectively accurate solution. This chapter will be concerned with the derivation
of a subjectively accurate image from such a solution.

The first sections of this chapter focus on obtaining an approximation from

the radiosity solution that is more suitable for rendering. These include ap-
proaches that construct a higher order approximation based on the radiosity
solution, as well as methods that partially reevaluate the radiosity equation at
image resolution. Hardware-assisted rendering and techniques for rendering tex-
ture and bump maps using radiosity are also described.

Section 9.5 examines the question of how to map radiance values resulting
from the radiosity analysis to pixel values in such a way as to produce a sub-
jectively accurate image. Typically, allowable pixel values are integers in the
0-255 range for each of the red, green, and blue color channels of the display
device. The conversion should compensate as far as possible for the limited
range of radiance values and color resolution offered by typical monitors, as
well as for perceptual effects.

The radiosity method has thus far been presented for a monochrome (i.e.,
single color band) solution. This chapter concludes with a description of the
issues relating to color perception and the selection of appropriate red, green, and
blue pixel values to represent the more general spectra of the visible wavelengths
of electromagnetic radiation.

Radiosity and Realistic Image Synthesis 243

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.1 RECONSTRUCTING THE RADIOSITY FUNCTION

Figure 9.1: Mapping a pixel location to a position or surface in the environ-

ment.

9.1 Reconstructing the Radiosity Function

Given an eye location and viewing frustum (see Figure 9.1) a function x(p)
can be constructed to map the center of a pixel (p = (px , py)) to a point xp ,
on a surface in the environment. This mapping is equivalent to tracing a ray
from the eye through the pixel until it hits a surface at point xp . The most
straightforward choice for the radiosity at this point is simply the value of the
approximate radiosity function B (xp ). This value can be found by evaluating
the basis functions that are nonzero at that location and summing the results,
weighted by the corresponding nodal values:
n

B(x p) = B N (x )
i=1
i i p (9.1)

Repeating this process for each pixel produces a discrete map (or image) of the
approximate radiosity function.1
Unfortunately, an approximation that provides sufficient accuracy during
the solution may not be adequate for rendering. Visual perception gives certain
features more weight and others less, with the result that quantitative accuracy
does not translate directly to perceived quality. One consequence is that the
constant elements often used during the radiosity solution cannot be used during
rendering, since discontinuities in value are immediately perceived as unrealistic

1
This description ignores the issue of correctly sampling the finite solid angle of the
pixel as seen from the eye so as to avoid aliasing (jaggies) in the image. Antialiasing
the removal of these sampling artifacts, is beyond the scope of this text, For the cur-
rent discussion, it will simply be assumed that one or more samples have been used to
reconstruct a scalar radiosity value per pixel.

Radiosity and Realistic Image Synthesis 244

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.2: A composition of constant and linear elements for rendering.

(see Figure 9.2). Because the eye is sensitive to first derivative discontinuities,
a C1 approximation (continuous in the first derivative) may also be desirable.
Elements providing the desired continuity could be used throughout the
radiosity solution, but if the solution is otherwise accurate enough, this unnec-
essarily increases its cost. Instead, most radiosity implementations derive a new
approximation with the desired characteristics just prior to rendering. Several
methods that take this approach are described in the following sections.

9.2 Interpolation Methods for Rendering

When the original solution approximation B is not adequate for rendering, it
can be replaced by another approximation using a new set of basis functions
and coefficients. The new basis functions are selected for

1. the desired degree of continuity to avoid perceptual artifacts,

2. fidelity to the original approximation,

3. and (if appropriate) suitability for fast rendering using hardware Gouraud
shading.

9.2.1 C0 Interpolation
If linear elements, which provide a C0 approximation, are used in the solution
they can be applied directly for rendering. However, the radiosity solution is
often performed with constant elements. In this case, a linear interpolation or
approximation must be derived from the constant element solution to produce C0
shading in the image. A piecewise linear interpolation can be constructed

Radiosity and Realistic Image Synthesis 245

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.3: Creating nodes at element corners using nodal averaging.

by creating linear elements with nodes at the element vertices. The values at
these vertices are then derived from the nodes of the original constant elements.
One approach to constructing a C0 interpolant is to create new nodes at
the corners of the existing elements. The radiosity at these nodes can then be
obtained by averaging the radiosities of the adjacent elements [62] (see Fig-
ure 9.3). Values at nodes on the boundary of the geometry can be extrapolated
from interior nodes, although care must be taken not to create negative values.
Finally, values in the interior of the elements are linearly (in triangular elements)
or bilinearly (in the case of quadrilateral elements) interpolated from the new
nodal values at the vertices. Nodal averaging tends to smooth the radiosity func-
tion, which can often help mask artifacts. On the other hand, smoothing may
eliminate subtle but important shading variations.
Another approach is to create a new element subdivision by connecting the
existing nodes at the centers of each original element with edges, thus using
the same set of nodes during the solution and rendering stages. For example,
to define constant elements for the solution, Sturzlinger [227] computes the
Voronoi diagram for the nodes and uses the Voronoi polygons as elements. For
linear interpolation during rendering, the corresponding Delaunay triangulation
provides elements with nodes at the vertices.
It is also possible to use a modified constant element with nodes located at
element corners instead of the center during the solution. Wallace et al. [247]
use this hybrid of constant and linear elements. During the solution, radiosities
are computed directly at the corner nodes. When needed, the constant radiosity
value for the element is taken as the average of the nodal values.2 The nodes

2
For very little extra effort, of course, these elements could be treated as true linear
elements during the solution [231].

Radiosity and Realistic Image Synthesis 246

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.4: Barycentric interpolation.

can then be used directly for bilinear interpolation during rendering. However,
evaluating the radiosity equation at element vertices has to be handled carefully.
Problems may occur where surfaces meet and the kernel of the integral operator
becomes singular, such as along the boundary of the floor where it meets a wall.

Bilinear Interpolation
Once values have been obtained at element comers, linear or bilinear interpo-
lation can be used to determine the radiosity at each pixel during rendering.
Since interpolation is performed in the parametric coordinates of the element,
the parametric coordinates of the point to be shaded are required. Depending
on the rendering algorithm, this may require transforming the world coordinates
(x, y, z) of the point xp , to the parametric coordinates of the element.
In the case of a triangle, barycentric coordinates provide a natural parame-
terization for linear interpolation. If the three vertices of a triangle are labeled
U, V, and W, with corresponding barycentric coordinates (1, 0, 0), (0, 1, 0), and
(0, 0, 1), then the barycentric coordinate u of a point in the triangle is simply
the ratio of the area Au of the subtriangle opposite the vertex U to the area of
the whole triangle (see Figure 9.4):

A = Au + Av + Aw

u = Au /A (9.2)
v = Av /A
w = Aw /A

The radiosity B(u, v, w), is then interpolated from the corner radiosities, Bu ,

Radiosity and Realistic Image Synthesis 247

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.5: Determining parametric (u, v) coordinates from (x, y, z) world co-
ordinates.

Bv , Bu , using

B(u, v, w) = u * Bu + u * Bv + w * Bw (9.3)

as shown in Figure 9.4.

For bilinear interpolation on planar quadrilateral elements, the transforma-
tion from world coordinates to the parametric coordinates of the element involves
solving a quadratic form that is guaranteed to have real roots for convex quadri-
laterals. The basic equations are described below. An efficient algorithm is
described in detail, along with several practical issues, by Haines [108].
The first step is to project the element and intersection point onto one of
the three orthogonal planes XY, XZ, or YZ. For numerical reasons, the plane
chosen should be the one corresponding to the largest component of the element
normal. For example, if the element normal is (0.10, -0.50, 0.86) then the
XY plane is best since the Z component of the normal is largest in absolute
value. The projection onto the XY plane then simply involves dropping the
Z coordinate of each vertex and the intersection point. (The equations below
continue with (x, y) but (x, z) or (y, z) can be substituted in a straightforward
manner.)
If the intersection point is (x, y) and the four corner vertices are (x00, y00)
for the vertex at (u, v) = (0, 0), (x01, y01) for the vertex where (u, v) = (0, 1),
etc., then (see Figure 9.5)

v = Z
x x +x Z x Z (9.4)
u = x00 + x00 Z 00x01Z 00+ x11Z +01x10 x10 Z

Radiosity and Realistic Image Synthesis 248

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.6: Bilinear interpolation.

where
Z = real solution to aZ 2 + bZ + c = 0
a = y 10x00 + y11x00 + y10x01 y01x10 y00x11
+ y00x10 y11x01 + y01x11
b = 2y10x00 + y01x10 2y00x10 + yx01 + yx10 (9.5)
y10x01 yx11 yx00 + y00x + y11x y01x
y10x + y00x11 y11x00
c = y10x y10x00 yx10 + y00x10 y00x + yx00

Once the (u, v) parametric coordinates of the point have been obtained, the
bilinear interpolation of the radiosities at the corners B00, B01, B11, B10 of a
quadrilateral element is

B(u, v) = (1 u)(1 v)B00 + (1 u)vB01 + uvB11 + u(1 v)B10 (9.6)

as shown in Figure 9.6.

Gouraud Shading
Since bilinear interpolation is performed in the parametric coordinates of the
element, it is independent of the projection of the element onto the image plane.
For reasons of computational speed, however, some algorithms perform linear
interpolation directly in image space. Gouraud shading, the most common exam-
ple, is an incremental formulation of linear interpolation designed for hardware
scan conversion [103]. During hardware rendering, colors defined at polygon
vertices are linearly interpolated along the polygon edges in the screen projec-
tion. At each scanline, these values are then interpolated horizontally between
the two edges intersecting the scanline (see Figure 9.7).
A hardware graphics accelerator can be used to render a radiosity solution
at high speed by passing the mesh elements to the hardware as polygons, with

Radiosity and Realistic Image Synthesis 249

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.7: Gouraud interpolation.

Figure 9.8. A polygon rendered at three different orientations using Gouraud

shading. The numbers indicate the color at the polygon vertices. The polygon is
rotated 45 degrees clockwise at each step moving from left to right. Courtesy of
Eric Haines, 3D/EYE, Inc.

Radiosity and Realistic Image Synthesis 250

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

the vertex colors determined by the nodal radiosities. Unfortunately, Gouraud

shading has limitations that can severely degrade image quality if not taken into
account. Since the perspective transformation from world space to image space
is nonaffine, linear interpolation in image space is not equivalent to interpolation
in the parametric space of the element.3 Furthermore, Gouraud shading does
not uniquely determine the interpolated value on the polygon interior for a given
location. Gouraud shading is dependent on the orientation of the polygon on
the image plane, and the interpolated shading may change visibly as the view
changes. This is particularly obvious in the case shown in Figure 9.7, where
for one orientation the value at point P depends on the vertices A, B, and C,
and in the other on B, C, and D. Figure 9.8 shows a polygon rendered at three
different orientations using Gouraud shading.
Since Gouraud shading interpolates only between the two closest edges
spanning the scanline, it also handles concave polygons poorly and will often
introduce discontinuities between neighboring scanlines on the polygon interior
(see Figure 6.11). A discontinuity at polygon boundaries is also possible if
adaptive subdivision is allowed to introduce T-vertices into the mesh. When the
view is oriented so that the collinear edges joined at the T-vertex fall along a
scanline, interpolation along the next scanline below or above will ignore the
T-vertex, and there will usually be a visible discontinuity in shading between
the two scanlines.
The limitations of Gouraud shading can be avoided with careful meshing.
For example, concave elements and T-vertices can be avoided using methods
described in Chapter 8. Baum et al. [18], for example, triangulate elements con-
taining T-vertices before rendering and thus avoid the view-dependent discon-
tinuities that can occur in such cases. Gouraud shading is orientation invariant
for triangles (except for the effect of the perspective transformation described
above), so triangulating all elements before sending them to the hardware will
largely eliminate view-dependent interpolation effects.
Since Gouraud shading is a form of linear interpolation it cannot be used
directly to render higher order elements if they have been used for the radiosity
solution. In such cases, hardware display can be accomplished by recursively
tessellating (subdividing) the elements into smaller polygons whose vertex values
are determined from the higher order interpolation. The subdivision ends when
some flatness criterion is fulfilled. The smaller polygons can then be passed to
the hardware for linear interpolation to provide an approximation of the original
element.

3
Since the perspective projection is nonlinear, distances across a polygon in the image
plane do not map linearly to distances measured on the actual polygon, Some hardware
implementations correct for the perspective transformation during interpolation.

Radiosity and Realistic Image Synthesis 251

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING

Figure 9.9: A comparison of C0 and C1 interpolation. The image on the left

was rendered using bilinear interpolation. The image on the right was rendered
using a C1 element constructed from quadratic triangular Bzier patches. Mach
bands corresponding to element boundaries are evident on the inside face of the
fireplace and along the ceiling panels. Courtesy of Mark Reichert, Program of
Computer Graphics, Cornell University.

9.2.2 C1 Interpolation
Although linear interpolation is C0 and is thus an improvement over constant
shading, it does not eliminate all perceptual problems. Linear interpolation is
still discontinuous in the first derivative at element boundaries, with the result
that Mach bands often appear along mesh boundaries (see Figure 6.2). A liner
mesh density or higher order interpolation, such as the quadratic interpolation
offered by some graphics hardware [139], can reduce or eliminate Mach bands,
since it tends to decrease the difference in slope across discontinuities. However,
piecewise quadratic or higher order interpolation does not necessarily provide
higher than C0 continuity at boundaries. Elements that provide a C1 approx-
imation will eliminate Mach bands in most cases without the need for a liner
subdivision.
The problem of interpolating or approximating the radiosity function dur-
ing rendering is analogous to the problem of constructing a surface in three-
dimensional space for modeling, where C1 or C2 surfaces are desired. The
surface modeling literature is thus a valuable source of methods relevant to
achieving approximations with particular continuity requirements. In the case of
radiosity, the surface to be interpolated is defined by the two parametric coor-
dinates of the geometry, with the third coordinate being the radiosity value (see
Figure 3.1 ).

Radiosity and Realistic Image Synthesis 252

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

A surface that is everywhere C1 can be constructed using quadratic or higher

order Bzier patches. A Bzier patch describes a surface that passes through
control points on the corners of the patch. Other control points on the boundary
and interior determine the tangent to the surface at the patch corners and the
shape of the surface on the interior. Max [161] has used cubic triangular Bzier
patches in screen space to produce C1 shading. Interpolation using quadratic
triangular Bzier patches has been applied to radiosity by Reichert [192], using
a scheme based on Powell and Sabin [183] and Cendes and Wong [45]. The
advantages of a C1 approximation for rendering are demonstrated in Figure 9.9,
in which the image on the right was computed using Reicherts approach. Note
the Mach bands corresponding to element boundaries in the left-hand image, for
which linear elements were used.
Surfaces with the desired continuity can also be constructed using more
general B-splines, in which case the surface will generally not pass through
the control points (although control points can be computed so as to make the
surface pass through a desired set of values.) There are many references on
surface construction and continuity [17, 80, 170] that provide fuller explanations
of these issues.

Clough-Tocher Interpolation
Salesin et al. [203] have applied the Clough-Tocher element [56] to rendering for
radiosity. This element is constructed from three cubic triangular Bzier patches
and provides a C1 interpolant at boundaries. Salesin et al. have developed
algorithms that take advantage of the large number of degrees of freedom in this
construction to relax selectively the continuity at boundaries corresponding to
actual discontinuities in the radiosity function. When discontinuity boundaries
have been identified (at some cost), it is desirable that they not be smoothed
over by the interpolation.
The Clough-Tocher construction begins by determining radiosity values and
normals (cross products of derivatives) to the radiosity surface at the vertices
of the triangular elements used during the solution. Salesin et al. begin with
triangular quadratic elements, which have nodes at the triangle vertices and at the
midpoint of each side, but quadratic elements are not essential to the algorithm.
The construction ends with each triangle subdivided into three new cubic Bzier
triangles. The additional control points introduced in the formulation allow for
explicit control of the continuity at element boundaries.
The Clough-Tocher construction proceeds as follows:

The first step in the algorithm is to determine normals to the radios-

ity function at each element vertex. If linear or constant elements have
been used during the solution, this can be accomplished by fitting a plane

Radiosity and Realistic Image Synthesis 253

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.10: Computing the normal to the radiosity surface at the vertices of
the Clough-Tocher element.

Figure 9.11: A cubic Bzier patch is defined by 10 control points.

through adjacent nodes. In the case of the quadratic elements used by

Salesin et al., the normals are computed using the values at the vertex and
midpoint nodes (see Figure 9.10).

The normals are computed by fitting parabolas along each edge of the
triangle through the three nodal values, one for each of the two endpoints
and for the midpoint of each edge. The cross product of the tangents of
the two parabolas at each vertex provides a normal vector. Since each
corner of the triangle may also be a vertex of other adjacent elements,
this process is repeated independently for each adjacent element, and the

Radiosity and Realistic Image Synthesis 254

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

Figure 9.12: Three Bzier patches are constructed in the subtriangles created
by connecting the centroid to the vertices. The control points in each shaded
region are constrained to be coplanar to ensure C1 continuity across element
boundaries.

resulting normals at the vertex are averaged.

The second step is to construct three cubic Bzier triangular patches within
each element. A Bzier patch is defined by 10 control points as shown in
Figure 9.11. The value at a point B(x) within the patch is determined by
the barycentric coordinates of the point and the values of the 10 control
points using,

3! i j k
B(x) = u v w bijk (9.7)
0i, j,k3
i! j!k!
i+ j+k=3
where (u, v, w ) is the barycentric coordinate of the point, and the bijk
are the control point values. The indices ijk for each of the 10 control
points are shown in Figure 9.11. Note that the Bzier surface interpolates
the values at the triangle vertices. For example, when (ijk) = (300), then
3!
(u, v, w) = (1, 0, 0), and B(x) = 3!1!1! 1 * 1 * 1 b300 = b300.
The Clough-Tocher construction involves dividing the element into three
cubic patches by connecting the element centroid to each vertex (see Fig-
ure 9.12). The 10 control points of each subelement provide the degrees
of freedom required to ensure continuity across the element boundaries
and also make it possible to relax the continuity conditions when an edge

Radiosity and Realistic Image Synthesis 255

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.2 INTERPOLATION METHODS FOR RENDERING

represents a shadow or penumbra boundary.

All four control points in each of the shaded wedges are constrained to be
coplanar. These constraints are determined as follows:

1. The control point values at the original triangle vertices (labeled a

in Figure 9.12) are taken directly from the original radiosity values.
2. The control points adjacent to the original triangle vertices (labeled
b in 9.12) are constrained to lie in the plane defined by the adjacent
triangle vertex and its normal.
3. Three of the wedges include one control point from an adjacent
element, (labeled c in Figure 9.12) thus ensuring element-to-element
continuity. The control points of these wedges are determined by the
two control points labeled b computed in step 2, and by the normal
to the wedge taken as the average of the normals at the adjacent
element vertices (labeled a).
4. This provides three of the four values for each wedge containing a
control point labeled d in the figure. Thus, the values of the d con-
trol points are determined by the plane through the three previously
determined points (labeled b and c). The value for the center point
e is then simply the average of the three points labeled d.

A representative cubic Bzier surface for the Clough-Tocher construction out-

lined above is shown in Figure 9.13.
The constraints on the control points can be relaxed when a boundary that
is C0 or C1 (discontinuous in value) is indicated. An element edge can be
lowered to C0 by relaxing constraint 3 for the wedge that spans the adjacent
elements. If the original vertices of the elements are also to be lowered to C0,
than there is no unique normal at the vertex and a different normal can be used
for each adjacent element. Care must be taken to ensure that the control points
along the element boundaries still provide C0 continuity between elements. This
is accomplished by adjusting the computed normals of adjacent wedges until
their respective components along the element are the same. The continuity of
the radiosity function across element boundaries can be further relaxed to create
discontinuities in value. In this case, multiple instances of the control points at
single location are created, thus decoupling neighboring elements completely.
Pseudocode can be found in [203].
In Figure 9.14 interpolation using linear (a), C1 quadratic Bzier (b), and
Clough-Tocher elements (c) is compared to a reference solution (d). The mesh
consists of 58 triangular elements and incorporates discontinuity boundaries. The
two C1 methods eliminate the inappropriate Mach bands apparent in the linear

Radiosity and Realistic Image Synthesis 256

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWOPASS METHODS

Figure 9.13: The Clough-Tocher element, showing the control net and the ra-
diosity surface produced. The shaded triangles indicate the control nodes that
must be kept coplanar to generate a C1 surface.

example. However, C1 quadratic Bzier interpolation also incorrectly eliminates

Mach bands corresponding to actual C0 boundaries along shadow edges. It also
tends to overshoot near the discontinuities, causing unexpectedly bright regions.
In the reference solution, the Mach bands along the shadow boundaries fade
away at the corners of the shadow. Salesin et al. prove that the radiosity at
the apex of two noncollinear C0 boundaries is actually C1. The Clough-Tocher
interpolant reproduces this result. This example is an excellent demonstration
of the subtleties involved in reproducing actual shading effects.
Elements that use a Bzier basis, such as the Clough-Toucher element, pro-
duce a surface that passes through the nodal values. Reconstruction can also
be performed by methods that approximate the surface without actually interpo-
lating the nodal values. For example, Metaxas and Milios [165] use thin-plate
splines to reconstruct images from sparse, noisy samples in image space, as
might be obtained by Monte Carlo ray tracing. In general, the field of scattered-
data interpolation and approximation represents a largely untapped source of
potentially useful techniques for radiosity and image synthesis [86].

9.3 Two-Pass Methods

With the exception of the importance-based approach discussed in Chapter 7,
the radiosity methods described so far have been inherently view-independent.
Decisions concerning the size and placement of elements have thus not taken into

Radiosity and Realistic Image Synthesis 257

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

(a) (b)

(c) (d)
1
Figure 9.14: Comparison of linear (a), C quadratic (b), and Clough-Tocher
(c) elements to reference solution (d). Courtesy of David Salesin, Program of
Computer Graphics, Cornell University.

account the projected size of the element in the image. If the mesh resolution
is too coarse for a particular view, small features may be absent or incorrectly
represented. However, increasing the mesh density during the solution increases
the solution cost with no guarantee that the higher resolution will be adequate
for any other view.
One approach to ensuring adequate detail is to incorporate view information
into the error metric used to guide adaptive meshing. The algorithms of Heckbert
[119] and Smits et al. [220] take this approach and have been discussed in
previous chapters. Discontinuity meshing can also help guarantee that important

Radiosity and Realistic Image Synthesis 258

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

shadow details are captured.

Another alternative is to evaluate the rendering equation directly at image
resolution, using the radiosity solution to provide the approximation of secondary
illumination [192, 198]. This is a two-pass method, in which the solution to
the transport equation is partially performed in a view-independent first pass
and completed for a particular view in an image-based second pass. Two-pass
methods are particularly useful in handling non-Lambertian reflection, as will
be described in the next chapter, but they are potentially useful whenever some
aspect of energy transport contributes a high spatial frequency to the image.

9.3.1 Evaluating the Radiosity Equation per Pixel

The radiosity for the surface seen at a particular pixel, p, is described by the
integral radiosity equation,

B(xp ) = E(xp ) + (xp ) s B(x)G(xp , x)dA (9.8)

where xp is the point in the environment visible at pixel p.

A radiosity solution provides an approximation B (x) that already takes into
account multiple reflections between surfaces. In the methods described so
far, this approximation has been used during rendering to determine radiances
directly by interpolation. An alternative approach is to substitute the approximate
radiosity solution B (x) for B(x) inside the integral on the right-hand side of
equation 9.8:

B (xp ) = E(xp ) + (xp ) s B (x)G(xp , x)dA (9.9)

Instead of using B (x) directly during rendering, this version of the integral is
then evaluated at each pixel. In effect, the final bounce of light rays to the
eye is computed at image resolution. This includes evaluating the geometric
factor G(xp , x) at each pixel, meaning that shadows are also resolved to image
resolution.
The decision to break the solution at this point may seem somewhat arbitrary,
but it is a reasonable choice, since higher order reflections tend to contribute
progressively less to the pixel variance [135]. High spatial frequencies in the
radiance of surfaces that illuminate a diffusely reflecting surface are averaged
during integration over the hemisphere of incoming directions. Intuitively, in-
accuracies in small, high-contrast features of the radiosity function affect image
quality much less dramatically in contributing to the illumination of other sur-
faces than when they are seen directly.

Radiosity and Realistic Image Synthesis 259

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

Figure 9.15: Monte Carlo ray tracing, using the radiosity approximation to
determine radiance for secondary rays.

More generally, the computation of B(xp ) using equation 9.9 can be re-
stricted to locations in the image where the visible element has been identified
during the solution as liable to be inaccurate, according to any of the error
metrics discussed in Chapters 6 and 7. In the hierarchical formulation of Chap-
ter 7, one might go a step farther and only reevaluate the integral for specific
interactions that involve partial visibility and a large amount of energy.
Evaluating the integral of equation 9.8 at each pixel requires considerable
computation since it involves form factor type computations from the point xp .
Most of the form factor algorithms discussed in Chapter 4 are applicable. Several
algorithms using this approach are described in the following sections.

Monte Carlo Ray Tracing per Pixel

Rushmeier [198] describes a two-pass approach in which the first pass consists
of a constant element radiosity solution. During the rendering pass the rendering
equation is evaluated using Monte Carlo path tracing [135], with paths limited
to one bounce (see Figure 9.15). Multiple rays are traced from the eye through
each pixel into the scene. From the intersection point of each ray, a single ray
is traced to a random point on a light source and a second ray is traced in a

Radiosity and Realistic Image Synthesis 260

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

Figure 9.16: An image computed using Monte Carlo path tracing, with the ra-
diosity solution supplying the radiance at secondary ray intersections. Courtesy
of Holly E. Rushmeier, Program of Computer Graphics, Cornell University.

random direction chosen with a probability proportional to the cosine of the

angle measured from the normal.
The radiance at the point intersected by the secondary ray is obtained from
the radiosity solution. If the secondary ray hits a light source, it must be counted
as making a contribution of zero, since the sources are already accounted for by
the ray to the light. The final pixel color resulting from n rays through the pixel
is then
n

(x ) (E(x ) + B(x
1 L )) j
i i i (9.10)
n
i=1

where the ith ray from the eye intersects the environment at xi , the secondary
ray from xi to a light intersects the source at x iL and the secondary ray reflected
randomly from xi intersects the environment at x ij .
Rushmeier points out that meshing algorithms can be greatly simplified
when the radiosity solution is no longer required to capture shading features like
shadows in great detail. As with all Monte Carlo solutions, noisy images will
occur if too few sample rays are used. Figure 9.16 shows an image computed
using Rushmeiers algorithm.

Radiosity and Realistic Image Synthesis 261

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

Figure 9.17: Computing form factors to every element at every pixel to compute
the final bounce of the radiosity equation.

Form Factors per Pixel

In a two-pass method described by Reichert [192], the radiosity at each pixel
is computed by evaluating a form factor to every element in the radiosity mesh
(see Figure 9.17). This is ultimately less efficient than Rushmeiers Monte
Carlo approach, since it computes a contribution from every element, regardless
of its contribution to the pixel variance. For the same reason, however, this
approach has the advantage of never missing an important source of illumina-
tion. Although expensive, it produces images of extremely high quality (see
Figure 9.18.)

Reichert notes that although meshing requirements for the radiosity solu-
tion are less rigorous when the final bounce is evaluated at image resolution,
sampling in image space can exacerbate certain types of errors. For example,
the illumination provided by constant elements can cause shading artifacts for
nearby receivers (see Figure 9.19 (a)). Similarly, coherent errors introduced by
the form factor evaluation used at each pixel may create highly visible patterns
in the image, as shown in Figure 9.19 (b), where form factors are computed with
point sampling. Monte Carlo integration of the form factor may be preferable
in this case since it will produce noise that is less objectionable than coherent
errors.

Radiosity and Realistic Image Synthesis 262

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

Figure 9.18: Image computed by evaluating form factors at each pixel to every
element. Courtesy of Mark Reichert, Program of Computer Graphics, Cornell
University.

Figure 9.19: (a) Artifacts caused by pixel-by-pixel form factor evaluation in

proximity to large discontinuities in B . The high gradients on the wall near
the floor correspond to the boundaries of constant elements on the floor. (b)
Artifacts caused by coherent errors in the form factor evaluation. Courtesy of
Mark Reichert, Program of Computer Graphics, Cornell University.

Radiosity and Realistic Image Synthesis 263

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

Figure 9.20: Shooting rays to compute the direct component of illumination dur-
ing rendering. The direct component is interpolated from the radiosity solution.

Direct Illumination per Pixel

It is not necessary in a two-pass method to integrate over all the elements

that provide illumination. Typically only a few such elements contribute to
high frequencies in the image. Integration can be limited to these elements,
with the contribution of the remainder interpolated from the radiosity solution.
Shirley [212] describes a two-pass method based on this approach in which
only the contribution of direct illumination (illumination from light emitters)
is recomputed by integration during rendering. Direct illumination typically
arrives from small relatively high energy sources and is thus more likely to
produce high gradients in the radiance function, which generally require fine
sampling for accurate approximation. Indirect illumination usually arrives from
relatively low energy sources distributed over a large solid angle and is thus less
likely to create high gradients.
Shirley computes the indirect component of illumination in a first pass con-
sisting of a modified progressive radiosity solution, during which the direct and
indirect components are stored separately. In the second pass the direct com-
ponent of the radiosity function is reevaluated at each image pixel using Monte
Carlo sampling of the light emitters (see Figure 9.20). This is added to the in-
direct component interpolated from the radiosity mesh. An example, which also
incorporates bump mapping and specular reflections, is shown in color plate 41.

Radiosity and Realistic Image Synthesis 264

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.3 TWO-PASS METHODS

Kok al. [141] describe a generalization of this approach in which illumination

due to the most significant secondary reflectors as well as the emitters is re-
computed during rendering, based on information gathered during the radiosity
pass.

9.3.2 Multi-Pass Methods

Although two-pass methods take advantage of the view to focus computational
effort, they have limitations. While two-pass approaches can account for high
frequencies across surfaces with respect to the final reflection of light to the
eye, they also assume that the global first pass radiosity solution is a sufficiently
accurate approximation of the secondary illumination for the particular view.
Two-pass methods provide no mechanism for recognizing when this assumption
is violated and, when it is, for refining the first-pass solution accordingly. In
splitting the solution into separate view-dependent and view-independent passes,
two-pass methods explicitly break the connection that would ultimately allow
automatic refinement of the solution to achieve an image of a specified accuracy.
The importance meshing algorithm of Smits et al. [220] addresses this problem
directly.
The multi-pass method described by Chen et al. [52] is a generalization of
the two-pass method that offers another approach to overcoming this limitation.
In the multi-pass algorithm, the radiosity solution is merely the initial step of a
progressive rendering process. Over a sequence of steps the radiosity approx-
imation is eventually replaced by a Monte Carlo solution at each pixel (with
specular reflection also accounted for in later passes). The use of multiple steps
allows finely tuned heuristics to be used to determine where and when to commit
computational resources in creating an increasingly accurate image. However,
care must be taken in designing a multi-pass method. In particular:

1. Redundant energy transport must be avoided. Each possible path for light
transfer must be accounted for only once. This may require removing part
of the radiance computed in an earlier pass before summing in a more
accurate estimate. Transport paths are discussed in more detail in the
Chapter 10.

2. Heuristics should be unbiased. In the limit, as more computational re-

sources are used, the solution should always converge to the true solution.
This is often a subtle and difficult aspect to assess.

Color plates 35-39 show the improvement of images with successive passes of
Chens algorithm.

Radiosity and Realistic Image Synthesis 265

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.4 INCORPORATING SURFACE DETAIL

9.4 Incorporating Surface Detail

In the discussion of the radiosity method so far it has been assumed that surface
properties such as reflectivity are constant over each surface, or at least over
the support of an individual basis function or element. This has allowed the
reflectivity term to be moved outside the integral during the computation of the
linear integral operator.
Surface detail is often simulated in image synthesis by using a texture map
to specify the variation over the surface of properties such as reflectivity, trans-
parency or the surface normal. Techniques for incorporating texture mapping
and bump mapping (mapping the surface normal) into radiosity algorithms are
described in the next sections.

9.4.1 Texture Mapping

When surface reflectivity is specified by a texture map, the resulting shading
typically contains very high spatial frequencies. These are difficult to represent
adequately with a finite element approximation. As an alternative, during the
radiosity solution the reflectivity r of a texture mapped surface can be approxi-
mated by the average texture reflectivity (the texture color), , over the support
of a basis function. This is a reasonable assumption in most cases, since high
frequency variations in radiance are averaged together during integration.4
During rendering, computing the radiosity at a particular point on a tex-
ture mapped surface requires knowing the incident energy (irradiance) at that
point. This is computed explicitly in the two-pass methods previously discussed,
and texture maps are handled trivially in such methods. However, even if the
radiosities are interpolated from the solution, it is still possible to extract an
approximation of the incident energy.
For elements on a texture mapped surface, the nodal radiosities will have
been computed using an average reflectivity, . When rendering the texture
mapped surface, the radiosity B (xp ) at any pixel can be modified to include the
reflectivity at xp specified by the texture map:

B(x , ) (x p )
, (x )) = B(x (9.11)

p p p

The radiosity is simply multiplied by the texture mapped reflectivity for the pixel
over the average reflectivity [61]. This effectively undoes the multiplication of
4
This assumption can be violated when a texture mapped surface containing large
scale variations in color is located near another surface, For example, a floor with a
black and white tile pattern may produce noticeable shading variations on the wall near
the floor due to light reflected from the tiles.

Radiosity and Realistic Image Synthesis 266

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.5 MAPPING RADIOSITIES TO PIXEL COLORS

the incident energy by the average reflectivity that was performed during the
solution and then multiplies the incident energy by the reflectivity (xp ) specified
for that location by the texture map. The paintings in Color Plates 10, 13, 17,
and 22 as well as the floor tiles in color plate 45 are texture maps incorporated
into the image using this technique.
It is also possible to combine texture mapping with radiosity using hardware
accelerators that support texture mapping. This is discussed in section 9.7,
where the details of rendering radiosity using hardware graphics accelerators are
described.

9.4.2 Bump Mapping

A bump map represents variations in the surface normal using an array of values
mapped to the surface, analogously to a texture map. Like texture maps, bump
maps generally produce shading with high spatial frequencies. Hence, by nature,
bump maps are best handled by per-pixel shaders. Two-pass methods handle
bump maps in a straightforward manner by using the perturbed normal at each
pixel during the view pass when computing the final radiosity. The image by
Shirley (color plate 41), includes a bump mapped brick pattern that was added
during the rendering pass [213].
It is also possible to support bump mapping when interpolation is used for
rendering. During the solution, instead of computing a single radiosity value
at each node, the variation of radiosity with surface orientation at the node is
sampled by computing independent radiosity values for several surface normals
distributed over the hemisphere. During rendering, the perturbed normal at a
particular pixel is determined from the bump map. For each of the element
nodes, the radiosity corresponding to that orientation is interpolated from the
hemisphere sampling, and the radiosity at the pixel is bilinearly interpolated
from these nodal values.
Chen [49] implements this algorithm by applying the hemicube gathering ap-
proach. A single hemicube is computed at each node. The form factors for each
sampled surface orientation are computed from this hemicube by resumming the
delta form factors, which are modified from the usual values to correspond to
the new surface orientation. Chen samples 16 surface orientations at each node.

9.5 Mapping Radiosities to Pixel Colors

Having finally computed a radiance at every pixel, the remaining rendering
step consists of assigning a corresponding frame buffer pixel value, typically
an integer in the range of 0 to 255. If color is desired, the radiosity solution
will have been computed at several wavelengths, and the mapping will include

Radiosity and Realistic Image Synthesis 267

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.5 MAPPING RADIOSITIES TO PIXEL COLORS

a transformation to the red-green-blue (RGB) color space of the monitor. The

monochrome issues will be addressed first, followed by a short discussion of
color.
The ultimate goal is to construct an image that creates the same sensation
in the viewer as would be experienced in viewing the real environment. As
discussed in the introductory chapter, there are many obstacles to realizing this
ideal, most of which are not unique to image synthesis. These include the
nonlinear relationship between voltage and luminance in the display device,
the restricted range of luminance values available on the display, limited color
gamut, the ambient lighting conditions under which the image is viewed, as well
as the basic limitations of representing three-dimensional reality as a projection
onto a two-dimensional image plane.5

9.5.1 Gamma Correction

The first difficulty encountered is that monitors, in general, do not provide a
linear relationship between the value specified at a frame buffer pixel (which
determines the voltage of the electron gun at that pixel) and the resulting screen
radiance. Rather, the radiance, I, is related to voltage, V, by [84],

I = kV (9.12)

The value of varies between monitors, but is usuaully about 2.4 0.2. Thus
I 1
V = ( ) (9.13)
k
Therefore, assuming that the target radiance is in the available dynamic range
of the monitor, a voltage must be selected from the available discrete voltages,
I 1
V j = round( ) (9.14)
k
The adjustments for the nonlinear relationship between voltage and radiance
through the use of the exponent, 1/, is called gamma correction.

9.5.2 Real-World Luminance to Pixel Luminance

A more challenging problem is the limited dynamic range of the display device.
Luminance values experienced in the real-world range from 105 cd/meter2 (star-
lit forest floor) to 105 cd/meter2 (sun reflected off snow). In contrast, a typical
CRT can display luminances in the range of only 1 to 100 cd/meter. It is there-
fore necessary to map the values produced by the radiosity simulation to the

5
These issues are covered in much greater detail in Halls monograph [114].

Radiosity and Realistic Image Synthesis 268

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.5 MAPPING RADIOSITIES TO PIXEL COLORS

range available on the CRT. The goal is to produce a subjective impression of

brightness in viewing the image on a CRT that is equivalent to that experienced
in viewing the real environment.
One simple approach is to map the luminance values in the radiosity solution
linearly to the luminance range of the monitor. Unfortunately, the only thing
visible in images produced using a linear mapping will usually be the light
source, since its luminance is typically several orders of magnitude greater than
that of any reflecting surface. Radiosity implementations often get around this
by mapping the highest reflected luminance (as opposed to the light sources) to
slightly less than the maximum pixel value and set the light sources themselves
to the maximum value. The slightly less ensures that light sources appear
brighter than any reflecting surface.
This mapping is completely arbitrary, and it is thus difficult to judge from
the image what the appearance of the real environment might be under equivalent
lighting conditions. Tumblin and Rushmeier [238] demonstrate this using the
example of a room illuminated by a light emitting with the power of a firefly
versus a room illuminated by a light with the same geometry, but with the power
of a searchlight. Because of the linearity of the integral operator in the rendering
or radiosity equations, scaling each resulting luminance range by the maximum
reflected luminance will produce identical images!
What is required is a tone reproduction operator, which will transform lu-
minances to frame buffer values in such a way that the perceived brightness
of the image equals that experienced by a hypothetical observer viewing a real
scene having the same luminances. Tumblin and Rushmeier derive such an op-
erator from simple models of the display and viewing processes. Their work is
a good example of how a model of perception might be incorporated into image
synthesis.
Figure 9.21 contains a diagram of the processes of perceiving real-world
and synthesized scenes. In the real-world, the luminance of the scene, Lrw, is
received by the eye and converted to a subjective perceived brightness, Brw.6
This is represented by the real world observer, which transforms luminance to
brightness under the conditions of viewing in the real world.
For a simulated scene, the computed luminance (which is assumed to closely
approximate the real-world luminance, Lrw), is mapped to a frame buffer pixel
value, V, (assumed here to be a number ranging from 0.0 to 1.0) by the tone re-
production operator, which is to be derived. The pixel value is then transformed
by the display operator to a displayed luminance, Ldisp. Finally, displayed lu-

6
Luminance measures light energy in terms of the sensitivity of a standard human eye,
and is computed by integrating the spectral radiance weighted by the luminous efficiency
curve for the eye over the visual spectrum (see Chapter 2). Brightness is a measure of
the subjective sensation created by light.

Radiosity and Realistic Image Synthesis 269

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.5 MAPPING RADIOSITIES TO PIXEL COLORS

Figure 9.21: The goal is to match the perception.

minance is, transformed to perceived brightness, Bdisp, through the process of

perception by the observer. This is represented by the display observer, which
transforms luminance to brightness under the conditions of viewing a CRT.
The goal is to derive a tone reproduction operator that makes Bdisp, as close
to Brw as possible. Such an operator can be constructed using the concatenation
of three operators:

1. the real-world observer, (i.e., the perceptual transformation from simulated

luminance, Lrw, to brightness, Brw),

2. the inverse of the display observer, (i.e., the transformation of Bdisp to

Ldisp ), and

3. the inverse of the display device operator, (i.e., the transformation from
display luminance, Ldisp, to the required pixel value, V).

Radiosity and Realistic Image Synthesis 270

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.5 MAPPING RADIOSITIES TO PIXEL COLORS

Applying these three operators in sequence transforms the simulated luminance

to a pixel value that can be sent to the frame buffer. The inverse operators
in this sequence effectively undo the subsequent application of the display and
observer operators in advance, thus making the net overall transformation from
simulated luminance to perceived brightness equivalent to applying only the real
world observer. Since the real world observer transforms Lrw to Brw, the result
is the desired match.
Tumblin and Rushmeier formulate the above real-world and display observer
operators using an observer model based on the work of Stevens [225]. Stevens
model expresses the relationship between luminance, Lin, and perceived bright-
ness B by
log10 B = (Lw) log10 (Lin) + (Lw) (9.15)

where and account for the observers adaptation to the overall image lumi-
nance, Lw. This equation provides the model for both the real-world observer
and the display observer, with different and in each case corresponding to the
adaptation levels for the different viewing conditions. The and parameters
are given by

(Lw) = 0.41 log10(Lw) + 2.92

(9.16)
(Lw) = 0.41 (log10 (Lw )) + (2.584 log10 Lw) + 2.0208
2

where Lw approximates the average overall luminance of the viewed real world
scene in one case and of the synthesized image in the other.
The display operator is expressed in terms of the contrast between a dis-
played luminance Ldisp and the maximum displayable luminance Ldmax:
Ldisp
= V + 1 (9.17)
Ldmax Cmax

where V is the pixel value in the frame buffer, and Cmax is the maximum
contrast between onscreen luminances. The in V is the gamma term for
video display described earlier. The 1/Cmax term accounts for the effect of
ambient light falling on the screen on image contrast. The Cmax term is defined
by the ratio of brightest to dimmest pixels, typically a value of about 35 : 1.
Combining equations 9.15 and 9.17 and inverting them where called for, a
single operator is derived that takes a luminance, Lrw , to a pixel value, V:
1/
L rw / disp ( rw disp )/ disp
V = rw 10 1 (9.18)
L
dmax Cmax

This relation maps the results of the global illumination simulation into a best
choice of pixel values, thus providing the last step in the image synthesis process.

Radiosity and Realistic Image Synthesis 271

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.5 MAPPING RADIOSITIES TO PIXEL COLORS

Figure 9.22. Images produced after accounting for brightnesses and viewer
adaptation based on work by Tumblin and Rushmeier. Light source intensities
range from 105 lamberts to 103 lamberts in geometrically increasing steps by
100. Coutesy of Jack Tumblin, Georgia Institute of Technology.

The advantage of using equation 9.18 is demonstrated in Figure 9.22 for a

model in which the light source ranges in intensity from 105 lamberts to 103
lamberts in geometrically increasing steps. Linear scaling would generate the
lower left image no matter what the light source energy. Using Tumblin and
Rushmeiers tone reproduction operator, the sequence successfully (and automat-
ically) reproduces the subjective impression of increasingly bright illumination.
If hardware rendering is used to create the image, it will obviously be
impossible to perform a sophisticated mapping to the radiosity at every pixel.
Instead, the nodal radiosities will have to be transformed before they are sent
to the hardware as polygon vertex colors. The result will differ somewhat from
the foregoing, since the pixel values in this case will be obtained by linear
interpolation from the mapped values, while the mapping itself is nonlinear.
As Tumblin and Rushmeier point out, the observer model upon which this
operator is based simplifies what is in reality a very complex and not completely
understood phenomenon. For example, color adaptation is ignored. There is
also the question of what adaptation luminance to use in computing the and
terms. Tumblin and Rushmeier use a single average value over the view or
screen, but further research might explore whether or not the adaptation level

Radiosity and Realistic Image Synthesis 272

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

should be considered constant over the field of view.

However, this work is an important step toward quantifying the perceptual
stage of the image synthesis process. The significance of such quantitative
models lies in their ultimate incorporation into the error metrics that guide the
solution refinement. Current metrics, for example, may force mesh refinement
in regions that are bright and washed out in the image, and thus contain little
visible detail. The goal of a perceptually based error metric would be to focus
computational effort on aspects of the solution only to the extent that they affect
image quality.

9.6 Color
The topic of color has been ignored in the previous chapters partly because it
is a complex topic in its own right, and partially because it is not integral to
the explanation of the radiosity method. Each of the issues addressed in the
earlier chapters is valid (with a few provisos7), for a full color world as well as
a monochrome one. The geometric quantities such as the form factor are inde-
pendent of material properties like color. The solution process is independent of
color in as much as a separate solution can be performed for each wavelength
or color band of interest.
A full description of color and color perception cannot be provided here.
Instead, interested readers are referred to a number of excellent sources for more
detail [156]. Valuable information from researchers in image synthesis can be
found in [114, 166].
Questions of immediate interest addressed in this section are,

What and how many wavelengths or color bands should be used in the
radiosity solution?

Can a single scalar value be used for error criteria, and if so, how is a
reasonable achromatic value derived from the color model?

How can one transform values between color models to finally derive
RGB pixel values?

The following sections outline properties of human color perception and

how these relate to the selection of color models. A color model provides a
framework to specify the color of light sources and reflective surfaces. Color
7
0ne assumption in this process is that light of one wavelength is not reflected at
another. With the exception of fluorescent materials, this assumption is not violated. It
has also been assumed that light that is absorbed is not reemitted in the visible spectrum.
This again is true at normal temperatures.

Radiosity and Realistic Image Synthesis 273

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

models are used at three stages in the radiosity simulation: (l) as part of the
description of the environment to be rendered, (2) during the radiosity solution,
and (3) during the creation of the final image on the output device. The same
or different color models may be used at each of these stages. Different color
models and means to transform values between them are discussed below.
The selection of a color model for the radiosity solution depends on the
required accuracy of the color reproduction and available information about
the lights and surfaces contained in the input to the simulation. For example, if
colors are input in terms of red, green, and blue components, then the simulation
can proceed by computing red, green, and blue radiosity values at each node. The
results can then be mapped directly to the RGB phosphors on a CRT based on
the specific properties of the monitor. However, this simple RGB color model
contains inherent limitations that make it difficult or impossible to reproduce
exactly the subtleties of color in the real-world. A fuller description of the
visible energy (light) at each point requires a specification of the energy at all
wavelengths in the visible spectrum. If a full (or sampled) spectrum of the
emitted energy of light sources and of the reflectivity of materials is available,
methodologies can be developed that lead to more accurate color simulations.

9.6.1 Human Vision and Color

The normal human eye is sensitive to electromagnetic radiation (light) at wave-
lengths between approximately 380 and 770 nanometers. The familiar rainbow
presents the spread of the visible spectrum from the short (blue) to the long (red)
wavelengths. Light leaving a surface and entering the eye will generally contain
some energy at all wavelengths in the visible spectrum. The relative amounts
of energy at each wavelength determine the color we see. For example, equal
amounts of energy at all wavelengths produce a sensation of white light.
It might at first seem that producing an accurate color image would require
reproducing the complete details of the energy across the visible spectrum, in
other words, the amount of energy at each wavelength at each point on the
environments surfaces.
Fortunately for the image synthesis process the human visual system greatly
simplifies the problem. Experiments have shown that very different energy
spectra can produce identical color sensations. Two different spectra that produce
the same sensation are called metamers. The reasons for this phenomenon and
the way in which this fact can be taken advantage of in the image synthesis
process are discussed below.

Radiosity and Realistic Image Synthesis 274

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.23: Luminous efficiency function.

Luminous Efficiency Function

The eye is not equally sensitive to light of different wavelengths even within the
visible portion of the electromagnetic spectrum. The luminous efficiency function
(see Figure 9.23) describes the eyes sensitivity to light of various wavelengths
and can be used to convert between radiance, which is independent of perceptual
phenomena, and the perceptual quantity, luminance.8
Thus, when working in an achromatic context (black and white), it is best
to multiply the energy spectrum by the luminous efficiency function to obtain a
scalar luminance value that can be displayed as a gray value or used as a deci-
sion variable for element subdivision. As will be discussed below, certain color
models contain one channel devoted to carrying achromatic luminance infor-
mation and are thus a convenient model for image synthesis. Scalar luminance
values can also be derived from sampled wavelength-based models by weighting
the energy at each sample wavelength by the corresponding value in luminous
efficiency function.

Color Sensitivity

Color perception results from receptors in the eye called cones. It has been
found that normal human eyes have only three types of cones, each with distinct
responses to light across the visible spectrum (see Figure 9.24). One type of cone
is sensitive primarily to short (S) wavelengths, one to medium (M) wavelengths,
and the other to longer (L) wavelengths. Color blindness is believed to be caused

8
The corresponding units of measure from the fields of radiometry and photometry
are discussed in Chapter 2.

Radiosity and Realistic Image Synthesis 275

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.24: Response of three color cones in eye (after Meyer, 1986).

by a deficiency in one or more of the three cone types. It is the response of

colorblind individuals to color that has provided much of the evidence of the
three distinct color cones.
The fact that color perception appears to be a function of only three stimulus
values provides the basis for metamerism. If two light spectra, when integrated
against the cone response functions, result in the same three values, then the
human eye is unable to distinguish between them. This is the root of metamerism
and is of great help in image synthesis since it provides a basis to produce a wide
range of color sensations by combining only a very few sources. In particular,
combinations of light produced by the red, green, and blue phosphors of the
typical CRT can recreate most possible color sensations.

9.6.2 Color Matching Functions and the CIE Chromaticity Diagram

Given the above observation, a three dimensional color space can be organized.
Experiments, as depicted in Figure 9.25, have been conducted in which an
observer is shown light from a monochromatic (single wavelength) test source,
and is simultaneously shown the combined light from three other monochromatic
sources or primaries chosen at short, medium, and long (s, m, l) wavelengths.
The observer is allowed to adjust the (s, m, l) primaries until there is a match.
One can then plot the resulting amounts of the three primaries as the test source
wavelength is varied across the visible spectrum resulting in three color matching
functions, (s(l), m(l), l(l)). Although most test sources can be matched by a
positive combination of the three adjustable sources, some cannot be matched.
In this case, the observer is allowed to move one of the three primaries to the left

Radiosity and Realistic Image Synthesis 276

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.25: Color matching experiment and resulting color matching functions.

side and add it to the test source. This quantity is then counted as a negative
amount of that primary.
The three resulting matching functions now provide all the information re-
quired to match not only the monochromatic test sources, but any combination
of energy across the spectrum. A general spectrum E(l) can be represented
with a linear combination of the primaries (s, m, l) given by

E() = S * s + M * m + L * l (9.19)

Radiosity and Realistic Image Synthesis 277

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.26: CIE standard observer matching functions.

where
S = E() s()d
M = E() m()d (9.20)
L = E() l()d

In fact, the three matching functions can be replaced by any three independent
linear combinations of the original three matching functions, resulting in a new
color space defined by new matching functions. These can be used in exactly
the same way as the results of the original matching experiment.
A commonly used set of matching functions was developed in 1931 by
the Commission Internationale de lclairage (CIE) based on the standard ob-
servers matching functions. The CIE functions, x(), y(), and z(), are
shown in Figure 9.26. These were chosen such that the XYZ color space con-
tains all possible spectra in the positive octant and y() corresponds to the
luminous efficiency function.
Figure 9.27 shows the CIE XYZ color space. A direction emanating from
the origin in this space represents all multiples of a particular linear combination
of the three matching functions. Within the limits discussed in section 9.5, the
points along a line from the origin will all produce the same color sensation at
different brightnesses. The horseshoe-shaped curve indicates the directions in
the space corresponding to the set of monochromatic sources (i.e., the rainbow)

Radiosity and Realistic Image Synthesis 278

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.27: The CIE XYZ color space with cone of realizable color.

from approximately 400 to 700 nanometers. Any point (X, Y, Z) lying within
this cone of realizable color represents some linear combination of visible light.9
The triangle (X + Y + Z = 1) is also shown. A chromaticity diagram
can be constructed by projecting the set of points on this triangle onto the XY
plane. A point on this projection (x, y) represents a vector (X, Y, Z) where
X
x = X+Y+Z
Y
(9.21)
y = X+Y+Z
Figure 9.28 shows this projection including the location of the red, green, and
blue phosphors of a typical CRT. All possible colors on a CRT (the monitors
gamut) include only linear combinations of the RGB phosphors, which explains
why not all colors can be reproduced. In particular, it is impossible to display
saturated yellow-green colors on a CRT.
The CRT is also constrained by the dynamic range of the phosphors, as
was described earlier. Figure 9.29 shows the CRTS RGB color space and its
transformation into the XYZ color space. All possible colors on a CRT thus lie
within the skewed cube shown.

9
A point outside the cone of realizable color simply does not exist, as it would require
a negative amount of light at some set of wavelengths.

Radiosity and Realistic Image Synthesis 279

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.28: The CIE XYZ chromaticity diagram.

The transformation between XYZ and RGB spaces depends on the specific
phosphors of the monitor in question. Details of how to measure the phosphors
can be found in [114]. The NTSC transformation for a generic or standard
monitor is given by
X 0.67 0.21 0.14 R
Y = 0.33 0.71 0.08 = G (9.22)
Z 0.00 0.08 0.78 B
The approximate inverse is
R 1.730 0.482 0.261 X
G = 0.814 1.652 0.023 = Y (9.23)

B 0.083 0.169 1.284 Z
Hall [114] provides an appendix with code for this and other transformations
between color spaces.

9.6.3 Color Spaces and Image Synthesis

Given an understanding of color perception and representation, one is left with
the question of how handle color in an image synthesis system. The above
sections have discussed color in terms of a number of color spaces, including

Radiosity and Realistic Image Synthesis 280

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.29: RGB cube and monitor gamut within the CIE XYZ color space.

Wavelength: the full visible spectrum includes an infinite number of in-

dividual wavelengths. However, a finite number of discrete wavelengths
can be used to define a finite dimensional color space.

RGB: the red, green, and blue phosphor values.

CIE XYZ: a standard color space based on color matching functions.

Other color spaces exist for a variety of reasons. The YIQ space is designed
primarily for television with the Y channel carrying luminance. A color space

Radiosity and Realistic Image Synthesis 281

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

based on cyan, magenta, and yellow (CMY) is used for printing since inks
subtract light. Thus, in this context, CMY is complementary to the RGB space.
Hue, saturation, and value (HSV) and hue, lightness, and saturation (HLS)
spaces are also used for their direct mapping to human subjective descriptions of
color. Other color systems have been developed to attempt to create linear color
spaces in a perceptual sense through nonlinear transformations of the earlier-
mentioned primaries. These additional color spaces will not be discussed here
as they are not generally used for image synthesis directly. However, many
computer-aided design systems use them. For each color space, a transformation
to the XYZ space can be found.
Any of the three color spaces can be used for radiosity computations. The
solution step, for example the Gauss-Seidel or Southwell iterations discussed in
Chapter 5 or the PushPull steps in the hierarchical solutions of Chapter 7,
must be repeated for each dimension (or channel) of the selected color space.10
Independent of the choice of color space, the values should be stored in a
floating point format or a large enough integer format to handle many orders of
magnitude. The reason for this lies in the nonlinear response of the eye to light.
Thus, the transformation to one-byte (0-255) phosphor values should only take
place at the final display stage.
Conceptually, using an RGB color space throughout the image synthesis
process is simplest and requires no intermediate processing after converting light
source and reflection spectra into RGB. In fact, many CAD modeling systems
only allow specification of color in terms of RGB. However, this immediately
restricts the possible colors for both lights and reflective surfaces to the monitors
gamut. In addition, accounting for differences from monitor to monitor is very
difficult to incorporate into such a system.
The limitations of the RGB space would argue for a display independent
color space such as the CIE XYZ space. An additional argument for such a
system as the CIE XYZ is that the Y channel can be used directly as a measure
of luminance and thus provides a simple criteria for error metrics in decisions
such as element subdivision. In fact, one might choose to perform all radiosity
computations only on the Y channel until element subdivision has completed.
The X and Z channels can then be processed based on the final element mesh.
However, any three-dimensional coordinate space requires an a priori integration
of the reflection and light source emission spectra. This can cause inaccuracies
as light from one wavelength will influence another through this prefiltering
operation.

10
It is worth repeating that the form factor computation, are independent of color and
thus only need to be computed once.

Radiosity and Realistic Image Synthesis 282

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.6 COLOR

Figure 9.30: Color computations from reflectionemission spectra to image.

9.6.4 Direct Use of Spectral Data

Meyer argues for the use of a set of samples at discrete wavelengths as the pri-
mary color space [166]. This involves selecting specific wavelengths at which to
sample the reflection and emission spectra, performing the radiosity solution at
each sample wavelength, and then reconstructing the spectrum or directly con-
verting them to the CIE XYZ (see Figure 9.30). The XYZ to RGB conversion
can then be done for display on a particular monitor. The number and wave-
lengths of the sample of the visible spectrum should be based on perceptual data.
The larger the number of sample wavelengths chosen to represent the reflectivity
and emission spectra, the closer the approximation. However, since each sample
wavelength requires a separate solution step, the larger the number of samples,
the higher the computational cost. After a careful study of experimental data
(see the experiment outlined in Chapter 11), Meyer concludes that four samples
can in most cases provide a good balance of cost and accuracy. In particular,
given a choice of only four sample wavelengths456.4, 490.9, 557.7, and 631.4
nanometerswere shown statistically to produce the most accurate simulations
when observers were asked to compare synthesized images of the Macbeth Col-
orChecker Charts with the real charts. The XYZ components are then found by

Radiosity and Realistic Image Synthesis 283

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.7 HARDWARE ACCELERATED RENDERING

weighting the energies at each wavelength, as follows:

E456.4
X 0.1986 0.0569 0.4934 0.4228 E
Y = 0.0034 0.1856 0.6770 0.1998 = E490.9 (9.24)
Z 0.9632 0.0931 0.0806 0.0791 557.7
E631.4

Light sources characterized by spectra with one or more narrow bands will cause
problems in systems that rely on discrete wavelength sampling; however, most
reflectors exhibit smooth reflection spectra. The details of the derivations and
experimentation in Meyers studies are not repeated here. A set of C code
implementations can be found in the appendices of Halls book [114].

9.7 Hardware Accelerated Rendering

9.7.1 Walkthroughs
If views of the radiosity solution can be rendered quickly enough, an interactive
walkthrough of the shaded environment is possible. Airey [5] reports that the
sensation of interaction requires at least six frames per second. Thus, radiosity
solutions are often rendered using hardware graphics accelerators, in spite of the
limitations of Gouraud shading discussed earlier. This section provides a short
discussion of some of the practical issues with the use of hardware graphics
accelerators for radiosity rendering.
The basic approach is to define a view camera, then pass each element in
the mesh to the graphics accelerator as a polygon with a color at each vertex
corresponding to the (scaled) nodal radiosity. Light sources are turned off during
the rendering, since the radiosity simulation has precomputed the shading. If the
use of mesh primitives (e.g., triangular strip, quadrilateral mesh or polyhedron)
is supported by the hardware, they can be used instead of individual polygons
to speed up rendering further. The basic flow of data to the graphics pipeline is
shown in Figure 9.31.
It is straightforward to add specular highlights during hardware rendering.
In this case, one or more light sources are turned on, approximating the po-
sitions of the light sources used during the solution. Specular colors and the
Phong coefficient are defined as appropriate as the elements are passed down the
pipeline. Where the original geometry was defined with vertex normals, these
should be interpolated to the nodes and passed along with the other vertex data
for each element. The diffuse color of all polygons should be set to zero since
the radiosities at each vertex provide the diffuse component. Depending on the
hardware shading equation, it may be necessary to turn on the ambient light
source so that the vertex colors are included in the shading equation.

Radiosity and Realistic Image Synthesis 284

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.7 HARDWARE ACCELERATED RENDERING

Figure 9.31: Rendering radiosity using a hardware graphics accelerator.

9.7.2 Hardware-Supported Texture Mapping

Some hardware graphics accelerators support texture mapping. During render-
ing, data describing the texture map is passed to the accelerator, followed by
the polygons to which it applies. The mapping of the texture to the surface is
often specified by supplying a texture coordinate, (u, v), at each polygon vertex.
During rendering, the u, v coordinates are interpolated to each scanline pixel
(typically using Gouraud interpolation). The u, v coordinate at the pixel is used
to look up the color defined by the texture map for that surface location. This
color is then incorporated into the hardware shading equation.
Depending on how the texture color is incorporated into the shading equa-

Radiosity and Realistic Image Synthesis 285

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.7 HARDWARE ACCELERATED RENDERING

tion, it can be possible to apply texture mapping to polygons that have been
shaded using radiosity. The goal is to have the shadows and other shading vari-
ations computed by radiosity appear on the texture mapped surface. For this to
work, the hardware shading equation must multiply the texture color at a pixel
by the color interpolated from the polygon vertices. The polygon vertex colors
can then be used to represent the incident energy at the element nodes, with the
texture color representing the reflectivity of the surface. As described in Chap-
ter 2, the incident energy at a node can be obtained by dividing the radiosity
at the node by the surface reflectivity used during solution (usually the average
color of the texture map). The product of the incident energy and the reflectivity
determined from the texture map then gives the reflected energy or radiosity at
the pixel.
If u, v texture coordinates are defined at the original polygon vertices, they
will have to be interpolated to the element nodes during meshing. During ren-
dering the vertex u, v coordinates and vertex colors corresponding to the incident
energy are then passed down to the hardware for each element.

9.7.3 Visibility Preprocessing

Even with hardware acceleration, an adequate frame rate may be unattainable
for models containing tens or hundreds of thousands of polygons, particularly
after the polygons have been meshed. Models of this size are not uncommon in
architectural applications.
Airey [5] proposes an approach to accelerating hardware rendering that is
particularly appropriate to building interiors, where only a fraction of the model
is potentially visible from any particular room. Airey uses a visibility preprocess
to produce candidate sets of the polygons potentially visible from each room.
A candidate set includes the polygons inside the room, as well as those visible
through portals (typically doorways) connecting the room with other rooms.
During rendering only the candidate set for the room containing the eye point
needs to be passed to the hardware renderer. The preprocess is simplified by
allowing the candidate list to overestimate the list of potentially visible polygons,
since the hardware renderer makes the ultimate determination of visibility at each
pixel. Aireys algorithm uses point sampling to determine the candidate list, and
thus may miss candidate polygons.
Teller describes an algorithm that can produce reliable candidate lists in two
dimensions [234] and Funkhouser et al. [89] discuss the use of this technique
to support walkthroughs of a model containing over 400,000 polygons. For the
three-dimensional case, Teller [233] gives an efficient algorithm to determine the
volume visible to an observer looking through a sequence of transparent convex
holes or portals connecting adjacent cells in a spatial subdivision. Only objects

Radiosity and Realistic Image Synthesis 286

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 9. RENDERING
9.7 HARDWARE ACCELERATED RENDERING

inside this volume are potentially visible to the observer. The details of this
algorithm are beyond the scope of this book. However, the reader is encouraged
to investigate this work as it introduces a number of concepts and techniques of
potential value to future research.
In addition to the development of candidate sets for visibility, interactive
rates can sometimes be maintained by displaying a lower detail environment. If
the mesh is stored hierarchically, a low-resolution version of the mesh can be
displayed while the view is changing rapidly, and then replaced with a high-
resolution version when the user rests at a certain view [5].

Radiosity and Realistic Image Synthesis 287

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS

Chapter 10

Extensions
Radiosity demonstrates the potential power of finite element methods for global
illumination calculations, at least in the case of environments consisting of Lam-
bertian diffuse reflectors. Given this success, it is natural to ask whether this
approach might be generalized to handle a wider variety of global illumination
phenomena.
In Chapter 2, the radiosity equation is derived from a general model of
light energy transport by restricting the problem in various ways. For example,
diffraction, polarization, and fluorescence are ignored, on the assumption that
these make only small, specialized contributions to everyday visual experience.
Light is assumed to move with infinite speed, so that the system is in a steady
state. Scattering and absorption by the transport medium (e.g., the air) are
disregarded. Most importantly, the directional dependency of the bidirectional
reflectance distribution function (BRDF) is eliminated by limiting the model to
Lambertian diffuse reflection.
Although computationally convenient, some of these assumptions are too
restrictive for general-purpose image synthesis. This chapter presents approaches
to lifting the restrictions to Lambertian diffuse reflection and nonparticipating
media. Specialized light emitters, such as point lights, spot lights, and sky or
natural light, are also discussed in the context of a radiosity solution.

10.1 Nondiffuse Light Sources

Perhaps the simplest extension to the basic radiosity method is to allow light
sources to emit with a non-Lambertian diffuse distribution. The simplicity of this
extension derives from the fact that lights are normally predefined. Lights are
also typically treated as emitters only (i.e., they do not reflect light). However,
difficulties are created by the variety of light sources in common use, each of
which requires subtly different handling.
Computer graphics applications use a variety of ad hoc and physically based
light sources. These include

Radiosity and Realistic Image Synthesis 289

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

Isotropic Point light: light emanates from a point with equal radiant in-
tensity in all directions. The flux density falls off according to 1/r2.

Parallel light: the light source is at infinity in a particular direction. Thus,

the flux density is constant. Direct sunlight can be approximated using a
parallel light.

Spot light: light emanates from a point with a variable intensity that falls
off from a maximum as the direction deviates from a given axis.

General luminaires: light emanates from a point or area according to a

general distribution function defined either by a goniometric diagram (see
Figure 10.1 (d)), often available from lighting manufacturers [3], or by an
analytic functional form.

Sky light: light emanates from a hemisphere representing the sky, possi-
bly accounting for weather conditions and solar position (but not the sun
itself).

General light emitters are discussed in the context of a ray tracing algorithm in
[243]. General luminaires and/or sky light have been incorporated into radiosity
applications in [74, 144, 176].
Although conceptually simple, the inclusion of more general light emitters
into a radiosity solution requires some care, particularly with regard to units
and normalization. In previous chapters, emission has been specified in units of
energy/unit area/unit time (or power/area). Since point and parallel sources have
no area, they will require different units. Normalization relates to the problem
of defining the total power of a spot (or more general) light independently of
the width of the spot beam or the shape of the intensity distribution.
The following discussion will assume constant basis functions, but the basic
concepts apply equally to higher order basis functions.

10.1.1 Form Factors to and from Light Sources

With the assumption of Lambertian diffuse area light sources, the rows and
columns corresponding to the light source in the approximate integral operator
K are derived in a similar fashion to entries for reflectors. However, light sources
are usually assumed not to be reflecting surfaces. Thus, if the ith element is
a light source, the ith row of the matrix, Ki,*, contains all zeros except for a
one on the diagonal (since i = 0). The entries of the ith column, K*,i will, in
general, not be zeros. These will be the terms responsible for shooting the light
to the receiving elements.

Radiosity and Realistic Image Synthesis 290

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

Point or parallel sources obscure the intuitive definition of the form factor
somewhat, since they have no area. For the same reason, units of power/area
have no meaning for point light sources and the total power or power per solid
angle (power/steradian) must be used instead. Using the reciprocity relationship

Fij Ai = Fji Aj (10.1)

the total contribution of a light source i to the radiosity of an element j is

Bj due to i = j Ei Ai Fij/Aj (10.2)

In general, the new light sources will be defined in terms of power, which is
equivalent to the factor EiAi, as opposed to the emitted radiosity Ei (power/area).
Including such a light source into the matrix formulation requires modifying the
row and column corresponding to the source. First, for a light i the ith column
of K must be divided by Ai to account for the fact that the lights contribution
as represented by those terms will be in units of power rather than radiosity. The
corresponding entry Bi in the vector of radiosities is now interpreted in units of
power, since it is the power of the light source (i.e., it is multiplied by Ai to
account for the division in the ith column). The entries Kji of the matrix were
originally given by
A
Kji = j Fji = j Fij i (10.3)
Aj

Although Ai is undefined in this case, the division can be performed symbolically

to obtain the new entry
Kji = j Fij/Aj (10.4)

which is computable, since Aj is not zero.

All entries in row i of the matrix are zero, since the light is not a reflector,
except for the diagonal term 1i Fii. The row is multiplied by Ai , leaving
a one on the diagonal. This also results in the Ei term now also being in units
of power (i.e., Ei Ai ) as desired.
An alternative to incorporating the light source into the matrix formulation is
to handle specialized light sources in a separate step prior to the actual radiosity
solution. In this approach, the row and column of the light source is removed,
and the contribution due to the light source is computed for every element (or
node) in the radiosity system, using the appropriate equation for that source.
The resulting element radiosities are then used as the emission (E) values for
the subsequent solution of the matrix equation.
In the following sections the factors Fij will be derived for the various types
of light sources.

Radiosity and Realistic Image Synthesis 291

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

(a) Diffuse Area power

Light area

(b) Point Light power

(c) Directional Light

power
perp. area

(d) General Luminaire power

steradian

(e) Spot Light power

steradian

(f) Sky Light power

steradian

Figure 10.1: Types of lights.

Radiosity and Realistic Image Synthesis 292

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

10.1.2 Point Lights

The inclusion of an isotropic point source (see Figure 10.1 (b)) emitting with
equal radiant intensity in all directions can be accomplished by a shift in units
and a modification to the form factor. As discussed in the previous section,
the source can be specified in terms of its total power. The fraction of energy
leaving the source located at xi and arriving at some element, j, is then

1 cos j V (x , x )dA
Fij =

Ai 4 r2 j i j (10.5)

The 1/4 term converts the total power to power per steradian, and the remainder
of the integrand is the visible solid angle subtended by element j.

10.1.3 Parallel Lights

Parallel lights (see Figure 10.1 (c)) can be thought of as a point source at a great
distance, or a very large source with light emanating in only a given direction,
r
. An obvious application is the modeling of direct sunlight. In this case, the
appropriate units are power per perpendicular area, that is, the amount of power
per unit area falling on a surface oriented to the light. In this case, the form
factor is simply the visible projected area of element j:

r
Fij = Ai cos j V( , x j )dA j (10.6)

r
where V( , xj ) is the visibility of the infinite source from a point xj on element
r
j in the inverse direction of the light. The function V( , xj ) equals one if a
r
ray from dAj in direction does not hit anything, and zero otherwise.

10.1.4 General Luminaires

A more general lamp or luminaire may be a point source or an area source and
may have an anisotropic intensity. Often, luminaire manufacturers will supply a
goniometric diagram that specifies the radiant or luminous intensity (defined in
section 2.4.5) of the source in candelas over a range of directions [3]. Standards
for such diagrams are prescribed by the IES [4]. The diagram includes the effect
of shadowing and reflection by the light fixture. The complete specification of a
general point light thus includes the lights position, orientation, and goniometric
diagram.
In Figure 10.1 (d) a typical goniometric diagram depicts two perpendicular
slices through the intensity distribution in polar coordinates. These coordinates
are with respect to a main orientation axis. More complicated goniometric

Radiosity and Realistic Image Synthesis 293

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

distribution are, unfortunately, difficult to specify and do not have a standard

form.
Expressing the geniometric distribution in terms of power per steradian, the
goniometric diagram can be reformulated as a maximum power per steradian
Imax scaled by a polar function ranging from 0 to 1. The polar scaling function
r r
S( ) is defined to takes a direction, , away from the source and returns a value
between 0 and 1.
r
Interpolation is, required to obtain a value S( ) from the goniometric dia-
gram for a direction that does not lie on either of the two perpendicular planes
defining the distribution. Langunou and Tellier [144] suggest the following
method of interpolating smoothly between the given goniometric slices:
r
1. Project the direction onto the two planes. For examples if the main axis
is in the +Z direction and the diagram depicts the XZ and YZ slices, then
r
the projection of an arbitrary vector = (x, y, z) yields the new vectors,
(x, 0, z) and (0, y, z), with angles x = atan2(x, z) and z = atan2(y, z)
off the Z axis.

2. Perform elliptic interpolation:

r
S( ) = Sx ( x ) cos 2 x + Sy ( y ) cos 2 y (10.7)

3. Finally, divide the result by the maximum, Imax.

The form factor from a point light i to an element j can now be derived.
Again, the form factor is proportional to the solid angle subtended by j from
r
the point of view of the light and is scaled at each dAj by S( ):
r cos j
Fij =
Ai
S( ) 2 V (x i , x j )dA j
r
(10.8)
r
where is a vector from the light sources to dAj.
For general area lights, the goniometric diagram must be converted to lumi-
nance by dividing by the projected area of the source. For example if the light
intensity, I, is, given in terms of candelas (cd), then the luminance (cd/m2) is
given by
Le ( ) = 1 I (10.9)
cos Ai

In this case, the form factor must be integrated over the area Ai of the light and
normalized by dividing by Ai,
r cos j

Fij = I S( ) 2 V (x i , x j )dA j (10.10)
Ai Ai Aj r

Radiosity and Realistic Image Synthesis 294

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

Figure 10.2: Geometry for skylight.

10.1.5 Spot Lights

Spot lights, as commonly defined for computer graphics, are a special case of
the general luminaire where the intensity distribution is defined implicitly by a
simple function (see Figure 10.1 (e)). The most common functional form is a
r
cosine of the angle away from the axis, raised to an exponent, S( ) = cosn.
As for the general luminaire above, if the spot light is specified by its maximum
power per steradian, Imax, in the direction of the axis, and the power of the
cosine distribution is n, then the form factor is given by
cos n i cos j
Fij =

Aj r2
V (x i , x j )dA j (10.11)

10.1.6 Sky Light

Illumination from the sky (as opposed to the sun) can be considered as light
emanating from a hemisphere of infinite radius (see Figure 10.1(f)). The appro-
priate units in this case are again power per solid angle (power/steradian), but in

Radiosity and Realistic Image Synthesis 295

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

this case the solid angle is not from the source but rather the solid angle to the
source (see Figure 10.2). This does not present a problem, due to the reciprocity
relationship.
The CIE1 provides a number of formulae for estimating the luminance of
a point P on the sky hemisphere, depending on cloud cover and sun position.
For a completely overcast sky the luminance is given by
(10.13)
L( ) = Lz 1 + 2 cos (10.12)
3

where Lz is the luminance at the zenith. In this simple empirical model, the sky
luminance is assumed uniform in a circle at any given height, so the luminance
is a function only of the angle between the zenith and the point P. The sky
in this model is brightest at the zenith and darkest near the horizon. The value
of Lz is itself a function of the height of the sun. It should be noted that this
model is generally not accurate for low-lying cloud cover.
For a clear sky, the CIE gives the following function:

3 0.32 sec
+ 0. 45 cos )(1 e
2
(0. 91 + 10 e )
L( , ) = Lz 3z0 2
(10.13)
0. 274(0. 91 + 10 e + 0. 45 cos z 0

where Lz and are as above, z0 is the angle between the zenith and the sun,
and the angle between the sun and P (see Figure l0.2). The angle can be
computed from the angle formed by the projections of the sun and P onto the
ground plane (see Figure 10.2), using cos = cosz0 cos + sin z0 sin sin
[1].
If the zenithal luminance Lz is converted to radiance Rz (see Chapter 2),
then the form factor term can again be derived. This requires an integration over
r
the sky dome hemisphere, , as well as over element j. S( ) is again defined
r
as the ratio of the radiance in direction to Rz (zenithal radiance).2 The form
factor to the sky is then given by:
r
S( ) cos j r
Fij =

Ai
2
V ( , dA j ) dA j d (10.14)

Takagi et al. [229] provide a valuable discussion of sky light in the context
of the photo-realistic rendering of automobiles in exterior scenes. Nishita and
Nakamae [176] discuss sky light specifically in the context of radiosity. In
particular, they address the issue of determining occlusion with respect to sky
light, as well as techniques for interiors that receive sky light through windows.
1
Commission
r Internationale de lclairage
2
S( ) may return a value greater than one near the sun in the clear sky model.

Radiosity and Realistic Image Synthesis 296

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

10.1.7 Normalization
The use of the standard Lambertian diffuse area sources requires the specification
of the source in terms of radiosity, or power/area. Thus, if the area of the light
source is changed and the radiosity is held fixed, the total power will change in
proportion to the area. Similarly, the above derivations of the form factors for
general luminaires and spot lights required the source to be defined in terms of
power/sr. As a result, if the maximum power/steradian of a spot light is held
fixed and the exponent n is allowed to vary, the total power of the light will fall
as n grows.
It is often desirable to specify an area source in terms of total power, thus
allowing the size of the diffuse source to vary without affecting emission. It
is also useful to have spot lights or more general luminaires specified in terms
of total emitted power, with the spotlight function or the goniometric diagram
defining only the relative intensity distribution.
r
This requires a normalization to replace the scaling function S( ) (just 1
for Lambertian sources) with a probability density function that by definition
integrates to 1 over the sphere for directional sources and over the area for area
sources. The advantage in this system is that as the area of a diffuse source or
the distribution of the spot light or general luminaire changes, the total amount
of energy emitted by the source remains constant. This provides a much more
intuitive system for modeling lights and determining their relative contributions
to the illumination of the environment. An additional advantage is that Monte
Carlo sampling, as described in Chapter 4, becomes straightforward.
Providing this normalization in source specification requires the derivation
of a normalization scaling factor based on the size and/or distribution of the
source.
Lambertian Diffuse Area Source: This is straightforward in the case of
the diffuse source. The source i can be specified in terms of total power, and
the scaling factor is simply 1/Ai.
Spot Light: In the case of the spot light, the normalization factor is deter-
mined by integrating the distribution function over the hemisphere:
2 /2

0 0
cos n sin d d (10.15)

Note that in polar coordinates, a differential element on the hemisphere is given

by sin d d. The above integral has an analytic solution:
/2
cos
n+1
= 1 (10.16)
n +1 0
n +1

Radiosity and Realistic Image Synthesis 297

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.1 EXTENSIONS

Thus the normalization factor is simply n + 1. In other words, to specify a 100

watt spotlight with a spot size defined by n = 30, the maximum power/sr in the
direction of the axis should be given as 100 (30 + 1) = 3100 watts/steradian. 3
General Luminaire: A similar result can be obtained from a general spatial
distribution by scaling the power by the reciprocal of the integral of the distri-
bution over the hemisphere. A nonanalytic distribution will require numerical
integration over the distribution in polar coordinates.

10.1.8 Light Source Data

Data for electrical light fixtures can be obtained from catalogs such as [3]. The
ies Lighting Handbook [2] is a good general resource for interpreting these
source. Directional data for light fixtures given by a goniometric (i.e., direc-
tional) diagrams are often available from luminaire manufacturers, but online
versions are not yet widely available.
The emission spectrum for a light source is determined primarily by the type
of bulb, (e.g., incandescent, low pressure sodium, etc.). Relative power spectra
for different types of lamps are given in several sources (e.g., [2, 127]). These
spectra may be characterized by smooth curves, as for incandescent lights, or
by narrow spikes, as for mercury lamps. Spectra characterized by spikes may
need to be filtered before use, depending on the color model adopted.
Smooth emission spectra are generally characterized as black body emitters
parameterized by temperature T. For a given temperature T, the blackbody
spectral radiance distribution is given by Plancks distribution:

I b ( ) = 2C1/[5 {exp(C2/T ) 1}]

C1 ~ 0.595 108Wm4/m2
C2 ~ 14388mK , in m , T in K (10.17)

The spectral distribution and luminance for natural (sky) light depends on
time of day, latitude and sky conditions (e.g., clear or over cast). The different
spectral values for direct (direct line to the sun) and indirect (from the hemisphere
of the sky) can be found in the [2] or [177]. A rough approximation of a (clear
sky is a blackbody at 15000K, and for an overcast sky, a blackbody at 6500K
The luminance of indirect natural light is generally in the range of 1000 to 5000
cd/m2. Direct sunlight is well represented spectrally by a blackbody at 5800K
with a magnitude of approximately 1300 W/m2.

3
Note that wattages given for light bulbs represent consumed power, not emitted light
energy. Most of the consumed power is converted to heat. A typical tungsten filament
converts only a small fraction (about 5 percent) of the consumed power to visible light.

Radiosity and Realistic Image Synthesis 298

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

10.2 Directional Reflection

As described in Chapter 2, the reflective behavior of a surface is described by
a bidirectional reflectance distribution function (BRDF) defined over the hemi-
sphere of directions above the surface. The BRDF represents the complex in-
teractions of incident light with the surface and has a complicated shape, in
general.
It is convenient to treat this complicated function as the sum of three com-
ponents: Lambertian (or ideal) diffuse, glossy, and ideal (or mirror) specular
[118] (shown in Figure 2.12). Radiosity is limited to BRDFs consisting only of
the Lambertian diffuse component. Models for radiosity thus consist entirely of
surfaces with matte finishes.
Since the non-Lambertian components of reflection play an important part
in everyday visual experience, radiosity images, although compelling, are often
not completely realistic. The absence of highlights (the glossy reflection of light
sources) not only reduces realism, but removes an important visual cue to shape
and curvature. The restriction to matte finishes is also a serious limitation for
design applications where the evaluation of surface appearance is important.
Before discussing methods to incorporate ideal specular and glossy reflection
into the radiosity solution, we will introduce the notion of transport paths and
a notation that will simplify the discussion and comparison of algorithms.

10.2.1 Classifying Transport Paths

Producing an image requires accounting (approximately) for all photons that
leave the light source and eventually enter the eye. The sequence of surface
interactions encountered by a photon on its way from the light to the eye de-
scribes a path.4 Global illumination algorithms can be characterized by which
paths they consider and how they determine them.
Kajiya first makes the connection between the Neumann expansion of the
rendering equation (equation 2.52) and the sequences of surface interactions
encountered during the propagation of light [135]. The rendering equation is an
integral equation that can be expressed in terms of an integral operator, K,

u = e + Ku (10.18)

In the case of the rendering equation, u corresponds to the radiance function and
e to the emission term. A solution to integral equations of this type (an exact
4
In practice, algorithms more typically take each path as representing a packet of
photons (a ray or beam). Each packet starts from the light carrying a certain power,
which is reduced at each interaction to account for absorption [214].

Radiosity and Realistic Image Synthesis 299

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

solution in the case of the rendering equation) is given by the Neumann series

u = Kie = e + Ke + K2e + K3e + ... (10.19)

Each term of the series corresponds to an additional application of the operator.

Kajiya points out the following physical interpretation of this series: each
application of the operator, K, corresponds to an additional surface interaction, or
bounce, along a path from the light to the eye. Thus, the term K2e accounts for
paths in which light reaches the eye via two bounces. This interpretation provides
a useful way of comparing illumination algorithms, which Kajiya undertakes for
several classic shading techniques.
The usefulness of such a comparison can be increased by formally including
the split of the BRDF into components, since algorithms can often be character-
ized by how they account for these components. Several authors have taken this
approach; the following discussion is based on Heckbert [120], who also intro-
duces the use of regular expressions to provide a compact notation for describing
paths.
For simplicity, the BRDF will be split into a Lambertian diffuse and a specu-
lar component, as in Heckberts presentation. The specular component subsumes
the glossy and ideal specular parts. The kernel of the rendering equation can then
be split into diffuse and specular components. Similarly, the integral operator
becomes the sum of diffuse and specular operators K = D + S.
Expanded in terms of these operators, the Neumann series becomes

u = e + (D + S)e + (D + S)2e + (D + S)3e + ...

= e + De + Se + DDe + DSe + SDe + SSe + ... (10.20)

Each term in this equation represents a subset of the infinitely many paths that
start at the light source and enter the eye from a particular direction. The first
term, e, is the path leading directly from the light to the eye. The second term,
De, consists of paths that start at the source and are diffusely reflected once
before entering the eye, while the third term, Se, represents paths that contain
one specular reflection, and so on. In all, light reaching the eye from a given
direction may have traveled from the source via paths that include any number
and combination of diffuse and specular interactions (see Figure 10.3). Since the
series is infinite all possible light paths are accounted for. Similarly, an algorithm
that accounts (correctly) for all paths provides a solution to the integral.
Heckbert suggests the use of regular expressions5 to simplify the description
of paths. With the addition of L to indicate emission from the light and E to
5
A regular expression describes a language consisting of words constructed from a

Radiosity and Realistic Image Synthesis 300

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.3: Some of the paths contributing to the radiance for a particular
direction leading to the eye.

Heckbert suggests the use of regular expressions5 to simplify the description

of paths. With the addition of L to indicate emission from the light and E to
indicate absorption by the eye, any particular type of path can be represented
textually by a string of symbols from the alphabet consisting of the letters
{L,D,S,E}. As an example, light arriving via one bounce from a diffuse
surface follows the path LDE, and that arriving by one bounce from a specular
surface follows the path LSE.
The regular expression L(D|S)*E expresses the set of all possible paths
for the two-component reflection model. This expression simply states that light
leaving the source and reaching the eye may follow a path containing any number
(including zero) of diffuse and specular reflections in any order. Algorithms can
be characterized compactly by expressions that describe the subset of all possible
paths that they attempt to account for.
The most common type of local illumination model, called the Utah approx-
imation by Kajiya and widely available in graphics hardware, is characterized
by the expression L(D|S)E. This expression contains the paths LDE (lo-

5
A regular expression describes a language consisting of words constructed from a
given alphabet. The expression formally expresses the set of words that can be realized
in the language and consists of a string of characters constructed from the alphabet plus
the superscripts * and +, the symbol 1, parentheses (), and the empty string . The
superscript * indicates any number (including 0) of repetitions of the superscripted term.
For example, D* indicates the set { , D, DD, DDD, DDDD, ....}. The superscript +
is similar to * but does not include the empty string (i.e., there must be at least one of
the superscripted term). The symbol | means OR, and the parentheses have the obvious
meaning.

Radiosity and Realistic Image Synthesis 301

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.4: Paths handled by the Utah approximation.

cal diffuse shading) and LSE (the Phong highlight), as shown in Figure 10.4.
In addition to ignoring paths involving multiple reflection, such models often
ignore shadowing, and thus include nonphysical paths. Classical ray tracing
handles the paths LDS*E | LS*E, but with some restrictions (see Figure 10.3).
The sequence LS in any path approximates glossy reflection using the Phong
model, but subsequent specular reflections ignore all but ideal specular reflec-
tion. Distribution ray tracing extends ray tracing to account for all paths in
LDS*E | LS*E. Radiosity is limited to the paths LD*E since in its traditional
form it cannot handle specular reflection (see Figure 10.6).

10.2.2 Tracing the Transport Paths

A complete global illumination algorithm must account for all paths that start
at the light and end at the eye. However, because of absorption and directional
reflection, all paths do not contribute equally to the radiance reaching the eye.
An efficient algorithm distributes computational resources among the paths ap-
propriately, expending the greatest effort on paths that contribute most to the
image.

Eye-Ray Tracing

For portions of paths consisting of the sequence S*E, the BRDF provides an a
priori basis for estimating the importance of various paths. The eye position and

Radiosity and Realistic Image Synthesis 302

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.5: Representative paths handled by classical ray tracing.

the pixel location determine the outgoing direction for the final S interaction,
and the shape of the BRDF provides a guide as to which incoming directions are
likely to contribute significantly.6 (In the case of ideal specular reflection, the
outgoing direction determines the only contributing incoming direction.) Work-
ing backwards from the eye, each subsequent S interaction along the path can be
treated similarly. This provides the basic strategy for Whitted-style ray tracing
and distribution ray tracing.
The utility of this strategy hinges on the ability to distinguish a limited
number of incoming directions as important. As the brdf becomes increas-
ingly diffuse, the relative importance of different incoming directions becomes
correspondingly closer to equal and an increasing number of incoming direc-
tions must be considered for an accurate estimate (see Figure 10.7). At some
point this strategy becomes prohibitively expensive. In such cases, most ray
tracing algorithms give up and determine importance based on the location of
the luminaires rather than the BRDF.

Light-Ray Tracing

Specular reflection also provides an a priori importance estimate when working

forwards from the light source, for portions of paths consisting of the sequence

6
A posteriori sampling is also useful, since the actual contribution coming from various
incoming directions cannot be estimated until some samples have been taken. The graph-
ics literature contains a great deal of discussion of such strategies [71, 135, 140, 147].

Radiosity and Realistic Image Synthesis 303

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.6: Paths handled by radiosity.

LS*. In this case, a path starting at the light intersects a surface, which deter-
mines an incoming direction, and the BRDF provides a basis for weighting the
importance of all possible outgoing directions. This is the motivation for light-
ray tracing, which is often used to find what are loosely referred to as caustics7.
Light-ray tracing for this purpose was first described by Arvo [12], although it
was applied to shading by Appel very early in computer graphics [8]. Many
subsequent algorithms have also used this approach [120, 214, 272]. As with
eye-ray tracing, this strategy loses its utility once a D interaction is encountered
(see Figure 10.7).

Bidirectional Ray Tracing

If there is only one D interaction, light-ray tracing and eye-ray tracing can be
combined to account for all paths by having them meet in the middle at the
diffuse surface (an approach often called bidirectional ray tracing). Light rays
that end on the diffuse surface typically have their power deposited in bins
into which the surface is subdivided. Eye rays ending on the diffuse surface
then interpolate the stored power to approximate contributions that would be
obtained by tracing the path farther (see Figure 10.8).

7
Although the term caustic has a very specific meaning in optics, in image synthesis
it has often been used more generally to refer to the illumination of diffuse surfaces by
light reflected from a specular surface. Mitchell and Hanrahan discuss caustics in detail
in [168].

Radiosity and Realistic Image Synthesis 304

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.7: Light-ray tracing and eye-ray tracing. Neither provide an advan-
tage for path segments containing a diffuse interaction.

Bidirectional ray tracing handles the paths LS*DS*E. Paths are effectively
broken at the diffuse surface by averaging together the effect of many paths
landing within a certain bin. This takes advantage of the fact that the ideal
diffuse component of radiance usually changes fairly slowly, a consequence of
the fact that incident illumination is integrated over the entire hemisphere of
incoming directionsessentially a filtering operation.

Radiosity and Transport Paths

In one sense, radiosity is simply a generalization of the bidirectional ray tracing

strategy to multiple diffuse interactions. Radiosity restricts the number of paths
to be considered by agreeing to minimize the error over a region (Galerkin
method) or at discrete points distributed over surfaces (point collocation). The
interactions between the regions or points are then computed using numerical
integration. In effect, the form factor between two elements (or basis functions)
averages together the effect of paths that directly join the two elements.
Conventional radiosity handles the paths LD*E. When combined with
eyeand light-ray tracing, it can be extended to the paths LS*(D*)S*E (see
Figure 10.9). In this case, light rays are traced until their power is deposited
on an ideal diffuse surface. A radiosity solution is then performed with initial
radiosities determined from the deposited power. Finally, eye-ray tracing is
used to render an image, tracing paths from the eye through (potentially) several

Radiosity and Realistic Image Synthesis 305

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.8: Bidirectional ray tracing with a single diffuse interaction.

specular interactions to reach a Lambertian diffuse surface.

The inner sequence D* in the above expression, which is handled by the

radiosity step, is incomplete, since it is only a subset of the correct sequence,
(D*S*D*)*.8 This is because radiosity does not account for light that trav-
els from one diffuse surface to another via specular reflection by intermediate
surfaces.

The next sections will present methods for accounting for these missing
paths in the radiosity solution. Such methods can be classified as either implicit
or explicit. In an explicit method, the global illumination problem is formulated
in terms of the radiance L(x, ). An approximation of L(x, ) is computed and
stored for all surfaces, including those with a BRDF containing glossy or ideal
specular components.

In an implicit method, the problem formulation remains the same as in the

radiosity method, with the effect of reflection from ideal specular or glossy
surfaces included only to the extent that they affect Lambertian diffuse surfaces.
This is accomplished by modifying coefficients of the operator K. The glossy
and ideal specular components are never actually computed or approximated.
8
The string LS*(D*S*D*)*S*E is equivalent to the string L(S|D)*E.

Radiosity and Realistic Image Synthesis 306

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.9: Radiosity with eye-ray tracing and light-ray tracing.

10.2.3 Implicit Methods

Extended Form Factors

The discretized radiosity equation expresses the radiosity for each element as a
linear sum of the radiosities of all other elements. Somewhat surprisingly, it is
possible to rewrite the equation, leaving out certain elements and still form a
correct solution for the remaining elements. The new system will be correct as
long as the coefficients of the radiosity equation, the form factors, account for
all paths between the elements of interest. This can be used to incorporate the
effect of glossy and ideal specular surfaces into the radiosity solution.
As an example, if there are four elements {a, b, c, d} (see Figure 10.10)
element b can be eliminated from the computation by computing extended form
factors from all other elements to all other elements via element b. In the case of
constant basis functions, the extended form factor Fdba represents the fraction of
energy leaving element d and arriving at element a after reflecting from element
b. If all surfaces are diffuse, then Fdba = Fdb * b * Fba. An equivalent system
of form factors can thus be constructed given all extended form factors with b
as the intermediate element, by

1. adding Fibj to the term Fij, for all pairs ij, and

2. eliminating the bth row and column from K.

The resulting system is equivalent to the original system, insofar as the included

Radiosity and Realistic Image Synthesis 307

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.10: Extended form factors.

elements are concerned. Eliminating element b is much like the elimination

involved in linear equation solvers such as Gaussian elimination.
The extended form factor is a natural generalization of the notion of a form
factor, which simply expresses the fraction of energy leaving one element that
will arrive at another element. Eliminated elements provide additional multi-
bounce paths along which energy can be transported between the remaining
elements. The additional paths are accounted for by increasing or extending the
form factors between the other elements.
When does it make sense to exclude certain surfaces from the solution?
Whenever it is difficult or expensive to store an approximation of the radiance
for those surfaces. This is certainly the case for surfaces with a highly directional
BRDF, since an adequate approximation requires storing a detailed directional
distribution at closely spaced nodes. It may also be useful to exclude procedu-
rally defined objects, like fractal surfaces, for which it may be impractical to
generate a mesh. When excluded surfaces are rendered in the final image, their
radiance is computed using another algorithm such as eye-ray tracing, with the
radiosity solution providing the approximation of the ideal diffuse component
where needed.
The basic theory for extended form factors was introduced to image synthesis
by Rushmeier [197, 201]. Rushmeier provides an algorithm for computing ex-

Radiosity and Realistic Image Synthesis 308

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

tended form factors for paths that contain a single ideal specular bounce (DSD)
from a planar surface (i.e., a mirror). With these limitations it is possible to
compute extended form factors using the hemicube algorithm. Mirrors are sim-
ply treated as openings into a virtual mirror world, which consists of the entire
scene flipped about the plane of the mirror. The form factor to an element in the
mirror world is added to the form factor to the element in the normal world, to
obtain the total extended form factor. Rushmeiers mirror form factor algorithm
provides the first pass of the two-pass method described by Wallace et al. [246].
Hall [114] provides pseudocode for the mirror form factor algorithm.
Malley, Sillion, and others [142, 157, 218] compute extended form factors
using recursive ray tracing. This allows multi-bounce DS*D paths to be fol-
lowed and BRDFs with a glossy reflection component to be included. It also
extends the method to nonplanar surfaces. In addition, ray tracing can be used to
compute extended form factors that account for transparent, refracting surfaces.
Shirley [212] has demonstrated the use of extended form factors to eliminate
a Lambertian diffuse surface from the solution. In this algorithm, the radiosity
of the excluded surface is computed during rendering using Monte Carlo ray
tracing.
Color plates 43 and 44 provide a comparison of images computed with and
without the use of extended form factors. Note the light missing from the top
of the table in front of the mirror in the before image.

10.2.4 Explicit Methods

Implicit methods, like extended form factors, do not provide an approximation
of the directional component of the radiance, which must be computed by other
means during rendering. By contrast, in an explicit method the more general ren-
r
dering equation is solved to produce an approximation of the radiance, L(x, ).
The approximation must thus represent the directional as well as spatial variation
of the radiance.

The Global Cube

One straightforward approach to approximating both the directional and spatial

variations of the radiance is to use two sets of basis functions. The directional
radiance distribution at each node can be approximated using one set of basis
functions defined over direction, with the spatial variation interpolated across
elements using a different set of basis functions.
Immel et al. [132] approximate the directional radiance distribution at each
node using a global cube, in which the radiance for discrete directions over finite
solid angles is approximated by a constant value. The directional discretization
is determined by uniformly subdividing the faces of a cube (see Figure 10.11).

Radiosity and Realistic Image Synthesis 309

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.11: Discretization of directions with the global cube.

This allows the surface visible in each direction to be determined quickly, as in

the hemicube algorithm. If the number of nodes in the environment is n and
the number of discrete directions represented on the global cube is m, there are
r
now n m unknowns representing the radiance, L(x, ), at n positions in each
of m directions.
The radiance for a given outgoing direction at a node can then be expressed
as a function of the radiance for every outgoing direction for every other node
in the environment.
r r
L(x, ) = E(x, ) +
x'
r r r r r r
( , x, ) k(x, , x, ) L(x, ) A (10.21)
r r
The k(x, , x, ) terms are similar to form factors in that they specify the
r
fraction of energy leaving point x in direction and arriving directly at x from

Radiosity and Realistic Image Synthesis 310

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.12: Rendering using the results of the global cube algorithm.

r
the opposite direction . Note that the reflectivity terms now also depend on
direction.
The resulting (n m)2 matrix can be formulated and solved as in the
conventional radiosity method. The matrix is very large but extremely sparse,
since a given node can interact with another node via only a single pair of
incoming and outgoing directions. Immel et al. anticipate progressive radiosity
by solving for the element radiosities in roughly the order of light propagation.
During rendering, Immel et al. obtain the radiance at a node for the view
direction by linearly interpolating between the sampled outgoing directions that
bound the view direction (see Figure 10.12). Radiances on the element interiors
are obtained by linear interpolation from the resulting nodal values. Hall [114]
provides pseudocode for the global cube algorithm, with discussion of some
related practical issues.
While the global cube algorithm approximates all paths in L(D|S)*E, the
algorithm runs into trouble for highly directional BRDFs, as shown in Fig-
ure 10.13. The fine details of the reflection in the floor of this image are poorly
approximated and display obvious artifacts. An accurate approximation of a
highly directional radiance distribution requires an extremely fine mesh, as well
as a high resolution global cube, which results in impractically high storage
requirements and solution times. Furthermore, increasing the cube resolution
everywhere is inefficient, since high resolution is principally required for por-
tions of paths ending at the eye.

Radiosity and Realistic Image Synthesis 311

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.13: An image produced using the global cube algorithm. Courtesy of
David Immel, Program of Computer Graphics, Cornell University.

Iterative Refinement

Shao et al. [210] address ideal specular and glossy reflection using the conven-
tional radiosity equation by changing the meaning of the form factor. As in
conventional radiosity, the form factor from a directionally reflecting surface Aj
to a second surface Ai is the fraction of the total energy leaving Aj that reached
Ai . For directional reflection, however, the amount of energy reaching Ai from
Aj will depend on the actual distribution of outgoing radiance for Aj . For a
glossy surface, this is a nonuniform distribution that depends on the BRDF and
the distribution of incoming energy. Thus, unlike a true form factor, Shaos
form factor is not purely geometric and cannot be computed a priori.
Shaos modified form factor is computed using the hemicube to determine
which elements are visible in all directions. The delta form factor for a receiv-
ing element seen through hemicube pixel p depends on the radiosity of other
elements contributing incoming energy through other hemicube pixels. The con-
tribution of the incoming energy for hemicube pixel q to the delta form factor
for pixel p is weighted using the BRDF for the element. Thus, the form factor to
a given receiving element will depend on the radiosity of elements contributing
incoming energy from other directions.
Clearly, Shaos form factor cannot be computed a priori, since it depends on
already knowing the element radiosities. Instead, he uses an iterative approach
to refine the estimate of the factors gradually. The initial step is to compute a
conventional radiosity solution, ignoring directional reflection. Form factories can
then be computed with respect to all directionally reflecting elements, using, the
initial radiosity solution to provide an estimate of the incoming distribution of
light for those elements. The radiosity matrix is reformulated and resolved with
the new form factors. The improved estimate can be used again to refine the

Radiosity and Realistic Image Synthesis 312

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

form factors further. To save time, the hemicubes are computed only once, and
then stored to be reused during later iterations.
The converged solution accounts for light arriving at diffuse surfaces via
directional reflection from other surfaces. Shaos approach is an explicit method,
in that approximations of the directional radiance distribution are computed when
needed during the solution. However, these approximations are not suitable for
rendering, for the same reasons noted for Immels algorithm, and distributed
ray tracing is instead used to determine the directional component at image
resolution.
Hall and Rushmeier [113] describe improvements to Shaos basic approach,
including adaptive subdivision for directionally reflecting surfaces and extensions
to progressive radiosity.

Spherical Harmonic Basis

In the global cube algorithm, the use of constant basis functions of a fixed width
to represent the directional radiance distribution means that a very large number
of directions must be evaluated and stored for each node to ensure that small
features are captured. Sillion et al. [217] observe that, just as with functions
defined over a surface, higher-order basis functions defined over the sphere of
directions should allow the distribution to be approximated with fewer samples.
Since the BRDF and the directional radiance distribution are functions defined
over the hemisphere, Sillion proposes the use of spherical harmonics as a basis.
The first three terms in the series are plotted in Figure 10.14.
Spherical harmonics are an infinite series, analogous to the Fourier series,
that can represent a large class of functions defined over a sphere. The first three
sets of spherical harmonics are shown in Figure 10.14. Slowly varying functions
can be approximated relatively accurately using only a small number of terms
from the series. Sillion et al. find that if the ideal specular term is removed, the
remaining radiance distribution due to glossy and Lambertian diffuse reflection
can usually be adequately approximated by computing coefficients for the first
100 or so terms of the series.
Representing the BRDF, as opposed to the directional radiance, is more
complex, since the outgoing distribution represented by the BRDF varies as a
function of incoming direction. Sillion et al. approximate the distribution for a
single incoming direction using approximately the first 80 spherical harmonics.
However, since this outgoing distribution changes as a function of the incom-
ing direction, the distributions must be represented for all incoming directions.
Assuming isotropic reflection, this variation depends only on a single value, the
angle between the incoming direction and the surface normal. The variation of
the distribution with this angle is interpolated using a one dimensional cubic

Radiosity and Realistic Image Synthesis 313

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.14: The first three spherical harmonics. Coutesy of Franois Sillion,
Cornell University Program of Computer Graphics.

spline to approximate the function describing the variation of each coefficient

with incoming direction.
Given this method for approximating the distribution, the solution process
itself is a straightforward generalization of a progressive radiosity approach. At
each solution step, the directional radiance distribution of every node is updated
based on the directional distribution of a selected shooting element. Sillion et
al. use a point collocation approach to approximate the spatial variation of the
radiance function. Thus, energy is transferred from elements to nodes. Each
node then has a directional radiance distribution represented by a finite number
of coefficients for the spherical harmonics.
The contribution of a shooting element with area A to the radiance distri-
bution, L(x, ), for a receiving node located at x is given by
r r cos cos r
L(x, ) = A' , ) r 2 r( , )V(x, x )dA
L( x (10.22)

where the dot represents the hemisphere of directions, L(x, ) is the radiance
r
distribution at x and ( , ) is the bidirectional reflectance function for incoming
r
energy from direction .
Numerical integration over the element i is performed, as in Wallace et
al.s ray-shooting form factor algorithm [247], by breaking the element into N

Radiosity and Realistic Image Synthesis 314

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

for (each delta area, Ai) {

r
Compute direction from x to sample point x in Ai ;
Shoot ray from x to x ;
if (no occlusion) {
r
Evaluate radiance L(x, ) leaving x in the direction toward x ;
Compute resulting incident energy flux at x;
Retrieve BRDF coefficients from B-spline approximation for
r
incoming direction to x ;
Scale BRDF by incident energy flux to obtain
contribution to outgoing radiance distribution L(x, ) ;
Add L to the nodes cumulative radiance distribution L(x, ) ;
}
}

Figure 10.15: Pseudocode for updating radiance at a point.

Figure 10.16: The summation of radiance distributions due to multiple contri-

butions from shooting elements. Coutesy of Franois Sillion, Cornell University
Program of Computer Graphics.

smaller delta areas, A. The resulting summation over the delta areas is
r cos X cos Xj r
N

L(x, ) =
j=1
L( x j , j )
r 2j
r( j , )V (x, x j )dA j (10.23)

The visibility of each delta area from a node i is determined by shooting a ray.
The steps for updating the radiance distribution at a node at x due to the current
source element i are given in Figure 10.15.
The addition of several contributions to a node's cumulative radiance distri-
bution is shown in Figure 10.16. Since the radiance distribution and the BRDF

Radiosity and Realistic Image Synthesis 315

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

are approximated by spherical harmonics, scaling, addition and other opera-

tions on these distributions are operations on the spherical harmonic coefficients
themselves. These are described in detail in [217].
Spherical harmonics are not appropriate for approximating ideal specular
reflection since the BRDF in this case is a Dirac delta function. In fact, ex-
plicit methods in general encounter difficulties for ideal or near-ideal specular
reflection, due to the expense of approximating the highly directional radiance
distribution. In the case of the ideal specular component, Sillion uses extended
form factors computed using ray casting (see section 10.2.3). During render-
ing, the ideal specular component is added using recursive eye-ray tracing. The
glossy reflection component is obtained from the stored directional radiance
distributions and interpolated linearly across elements.
Images computed using Sillions algorithm are shown in color plates 46 and
49. Note the diffused highlight on the cabinet doors of the kitchen in color plate
49. In color plate 46 the object on the left has an aluminum finish. The bright
illumination in the upper left-hand corner of the containing box is due to light
bounced directionally from the top of the aluminum object.
In addition to spherical harmonics and Immels global cube data structure,
several other schemes for storing the directional radiance distribution have been
proposed. La Saec and Schlick [148] use a hemisphere discretized along merid-
ians and parallels. For a progressive radiosity approach, Chen [49] stores the
energy incident from each shooting source in a queue. The radiance distribution
is not actually resolved until the receiving element is called on to shoot its en-
ergy in the progressive solution. Buckalew and Fussell [40] describe a system
in which a network of links is constructed along parallel rays in space, for a
distribution of directions. Energy is then transferred along these links based on
the reflection of the intersected surfaces.

10.2.5 Non-Lambertian Reflection and Hierarchical Methods

Aupperle and Hanrahan [14] have made a thorough theoretical examination of the
issues related to the use of notions similar to extended form factors. Significantly,
they have also included these notions into the framework of hierarchical radiosity
methods.
In their work, the radiance distribution is not represented in terms of po-
sition and direction, but rather through two positions The second position is
interpreted as in the direction of. In other words, Lij specifies the radiance
of area i in the direction of area j. Note that there is no explicit specification
of direction; instead the direction is defined by the relative positions of areas i
and j. This redefinition of the radiance distribution provides the framework for
applying hierarchical methods similar to those outlined in Chapter 7.

Radiosity and Realistic Image Synthesis 316

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Constant basis functions can be defined over any (area area). In other
words, the radiance from area i to area j is approximated as constant over the
two areas. Given this approximation of the radiance field, one can derive the
relation,
L jk = E jk + Lij Rkji (10.24)
i
where, Ejk is the emitted radiance from area j to area k and Rkji is the area
reflectance of area j in direction k due to light from area i. Putting the definition
of area reflectance aside for a moment, the above relation can be stated as: The
radiance of area j in the direction of area k is equal to the emitted radiance
of area j in the k direction plus the sum of the radiance of ALL areas i in the
direction of area j multiplied times the area reflectance Rkji.
Equation 10.24 is analogous to the original radiosity equation for diffuse sur-
faces. The biggest difference is the area reflectance term, Rkji, which accounts
for the BRDF as well as the geometric and visibility relationships between areas
j and k and between areas j and i. The derivation of this term is not pursued
here; the reader is referred to [14]. The result is given as

Ai Aj Ak
f r (x, x , x )G(x, x )G(x x )dx dx dx
Rijk =
Ai Aj
G(x, x )dx dx (10.25)

where fr is the bidirectional reflectance distribution function and G is a geomet-

ric term similar to the form factor (without the normalizing in the denominator).
One can now ask what the optimal choices of areas i, j, and k are in
order to have as few Rkji terms as possible, while maintaining a given level
of accuracy. Aupperle and Hanrahan develop a hierarchical algorithm based on
arguments similar to those underlying the hierarchical form factor algorithm of
Hanrahan et al. (described in Chapter 7) [116]. The resulting algorithm can be
found in [14].

10.2.6 Transmission

Although only reflection has been discussed so far, surfaces may also transmit
light. Transmission can be described by a bidirectional transmission distribu-
tion function (BTDF), analogous to the BRDF for reflection, defined over the
hemisphere on the reverse side of the surface. As in the case of reflection,
transmission can be separated into Lambertian diffuse (translucency), glossy,
and ideal specular parts.
Note that the BTDF describes the interaction only at the interface between
the two transport media (for example, the air and glass). It does not take
into account absorption within the transmitting material, which depends on the

Radiosity and Realistic Image Synthesis 317

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

length of the transmission path. Absorption of a fixed fraction at the interface

is a reasonable approximation for thin transmitters like windows, but in general,
a correct treatment of absorption requires an illumination model that includes
participating media.
Transmission can be incorporated into a radiosity solution using implicit
or explicit methods, just as for reflection. In an implicit method, for example,
refraction could be included by using recursive ray tracing to compute extended
form factors. In explicit methods, which construct an approximation of the
reflected radiance distribution, transmission can be incorporated by constructing
a similar approximation of the transmitted radiance distribution.

Translucency

Translucency, or Lambertian diffuse transmission, is analogous to Lambertian

diffuse reflection and is particularly easy to add to a conventional radiosity so-
lution [197, 201]. In translucency, a fraction of the energy incident on one side
of a surface is reradiated by the other side of the surface in a Lambertian diffuse
distribution. Thus, energy incident on the back of a translucent element con-
tributes to the radiosity of the front of the element (see Figure 10.17). Likewise,
the back of a translucent element will have a radiosity that is partially due to
energy incident on the front. Since the transmission distribution is constant for
translucency, it can be represented by a single scalar value, just as for Lam-
bertian diffuse reflection. Only two modifications to the radiosity algorithm are
required.
First, the radiosity equation for a translucent element now includes a con-
tribution due to light arriving on the back side of the element,
N
Bi = Ei + ( B F
j=1
i j ij + i B j Fkj ) (10.26)

where i is the coefficient of Lambertian diffuse reflection, i is the coefficient

of Lambertian diffuse transmission, element k is the back side of element i. and
Fkj is thus the form factor from the back of element i to element j. The second
change is to recognize that there are now more elements since the front and
back sides of a surface may act independently. Thus n is the total number of
elements, counting the back sides of translucent surfaces as separate elements.
Aside from computing the extra radiosities and form factors, the radiosity
solution proceeds normally. The image in Figure 10.18 was computed using
translucent elements to model the lampshade. The image was rendered us-
ing Monte Carlo eye-ray tracing to compute the direct illumination (see sec-
tion 9.3.1).

Radiosity and Realistic Image Synthesis 318

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.17: A fraction of the light arriving at the front of a translucent surface
is transmitted and radiated from the back of the element with an ideal diffuse
distribution. Similarly, a fraction of light arriving at the back of a translucent
element is radiated from the front.

10.2.7 Two-Pass Methods

It is possible to account for all paths in L(S|D)*E using only eye-ray tracing
or only a finite element approach. Kajiyas path tracing algorithm provides an
example of the former approach, while Immels global cube radiosity algorithm
is an example of the latter. Neither is entirely satisfactory. Eye-ray tracing has
the advantage that the number of paths to be traced is reduced based on the
eye direction. Directional reflection can be sampled as finely as desired with no
storage penalty. However, for diffuse reflection, a large number of paths must
be sampled, and images tend to be noisy (see the Monte Carlo image in color
plate 1, for example).
The finite element approach reduces the number of paths that must be traced
for diffuse reflection by averaging the effect of paths over the surface. How-
ever, this approach requires excessive storage for highly directional reflection,
as demonstrated by [132] (see Figure 10.13).
Two-pass methods abandon strict adherence to a single strategy; instead,
they use each strategy to handle the category of paths for which it is most

Radiosity and Realistic Image Synthesis 319

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.18: Image computed using translucent elements to model the lamp-
shade. Courtesy of Peter Shirley, University of Indiana.

efficient. Because of the linearity of light transport as modeled by the rendering

equation it is possible to sum the independently computed components to get
the final result. Such approaches are often called two-pass methods, because
the total solution typically consists of a radiosity and/or light-ray tracing pass
followed by an eye-ray tracing pass.
The two-pass approach has been used prior to its application to radiosity
(and prior to the origin of the term two-pass). For example, Kajiya and Von
Herzen L 136] use a two-pass method to compute images of clouds and other vol-
ume densities. Other examples of nonradiosity-based two-pass methods include
Arvos backwards ray tracing algorithm [12], Heckbert and Hanrahans beam
tracing algorithm [124], and Chattopadhyay and Fujimotos bidirectional ray
tracing algorithm [48]. Watt [259] also describes a two-pass method in which
caustics are rendered by tracing beams forward from the light source.

Radiosity Plus Reflection Frustum

A two-pass method incorporating radiosity is first described by Wallace et al.

[246]. The first pass consists of a Gauss-Seidel radiosity solution with stored
form factors. Extended form factors are computed using Rushmeiers hemicube
based mirror form factor method. The second pass traces eye-rays by recur-

Radiosity and Realistic Image Synthesis 320

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.19: Use of the reflection frustum during the rendering pass to sample
a highly directional BRDF.

sive application of a Z-buffer-based sampling algorithm, called the reflection

frustum. The reflection frustum algorithm samples a number of incoming di-
rections using scan conversion and Z-buffer hidden surface removal to evaluate
the glossy component of reflection (see Figure 10.19). The sample directions
are selected as grid points on the cross section of a square frustum oriented in
the specular direction. The contribution of each sample is then weighted by the
non-Lambertian components of the BRDF in that direction. This combination
of algorithms accounts for the paths L(D*|D*SD*)S*E, with the restriction to
ideal specular reflection in the D*SD* sequences (which shows how difficult
it can get to figure out which paths are actually accounted for in any particular
algorithm).

Radiosity Plus Ray Tracing

The two-pass algorithm described by Sillion and Puech [218] differs from that
of Wallace et al. in that recursive ray tracing is used to compute extended form
factors in the first pass, which allows the inclusion of DS*D sequences (see
Figure 10.20). Eye-ray tracing is used to handle ideal specular reflection in the
second pass. Sillion and Puech point out that distribution ray tracing could be
used to include paths containing glossy reflection as well.
The algorithm of Sillion et al. [217] (described earlier in section 10.2.4)
in which glossy reflection is approximated using spherical harmonics, is also

Radiosity and Realistic Image Synthesis 321

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.20: Use of recursive ray tracing to compute extended form factors dur-
ing the first pass of Sillions two pass algorithm. Sillion uses a single projection
plane rather than the five-faced hemicube to compute form factors.

incorporated into a two-pass method. In this case completing the solution re-
quires only the use of eye-ray tracing in the second pass to handle S*E path
segments consisting of ideal Specular interactions. This algorithm accounts (ap-
proximately) for all paths in L(S|D)*E.

Bidirectional Ray Tracing

Bidirectional ray tracing, described above in section 10.2.4) is a two-pass method,

although as originally formulated it traces power from the light only until it lands
on a Lambertian diffuse surface. However, it can be extended to handle mult-
iple Lambertian diffuse interactions, using an approach similar to progressive
radiosity in which light rays are propagated further into the environment through
successive bounces from diffuse surfaces [213, 214].
Although equivalent to progressive radiosity, this approach does not explic-
itly compute form factors. Instead, power-carrying light rays are shot out into
the scene from the light emitter. When a ray hits a surface its power is deposited
and stored in an array of bins (analogous to elements). Light rays are then shot
in a cosine distribution from the diffuse surface reflecting the most energy, and
so on, for each reflecting surface (see Figure 10.21). As for progressive radios-
ity, the shooting step is repeated for other reflecting surfaces until the energy
falls below a certain level.

Radiosity and Realistic Image Synthesis 322

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

Figure 10.21: Light-ray tracing extended to handle diffuse interreflection. These

are two steps from the first pass of a two-pass algorithm.

Glossy and ideal specular reflections can be incorporated into the first pass
using an approach analogous to extended form factors. Instead of computing
and storing the radiance for directionally reflecting surfaces, rays that encounter
specular surfaces are traced recursively until they reach a diffuse surface. During
the second pass, eye-rays are traced to account for segments of paths that start
at a diffuse surface and interact with specular surfaces before reaching the eye
(DS*E path segments).
Color plate 42 was produced by Shirley using this approach. Note the caus-
tics on the table top, which are produced by S*DE path segments refracted
through the wine glass. These are computed during the light-ray pass. The re-
fractions and reflections visible in the wine glass itself are DS*E path segments
followed during the eye-ray pass. In addition to tracing eye-rays to follow spec-
ular reflection, Shirley also computes LDE paths, i.e., the direct illumination
of diffuse surfaces, at each pixel during rendering (this approach is discussed in
greater detail in Chapter 9). Shirley uses Arvos illumination map data structure
[12] to store incident power on diffuse surfaces.
Heckberts rex algorithm uses the same basic light-ray tracing strategy as
Shirleys algorithm, as well as a more powerful adaptive version of the illumina-
tion map data structure (this algorithm is described in more detail in Chapter 8).
The images in Figures 8.30 and 8.31 were computed using Heckberts bidirec-
tional ray tracing algorithm.
The multi-pass algorithm of Chen et al. [52] uses both extended form factors
and light-ray tracing. A radiosity solution using extended form factors is first
performed to provide an initial low-frequency approximation of all L(D|S)*E

Radiosity and Realistic Image Synthesis 323

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.2 DIRECTIONAL REFLECTION

paths. Light-ray tracing is then used to produce a more refined high-frequency

approximation of caustics produced by LS*D path segments. A key point, men-
tioned by Chen, is that the contribution of LS*D path, initially computed using
extended form factors must be subtracted before the higher-quality approxima-
tion of the same paths made by light-ray tracing is added in. When combining
algorithms in two-pass or multi-pass approaches, care must be taken to avoid
counting the same paths more than once.

Radiosity Plus Phong

Where absolute accuracy is not an issue, satisfactory images can be obtained

relatively quickly by adding specular highlights to the radiosity shading dur-
ing hardware rendering. This is accomplished by defining light sources for the
graphics hardware prior to sending down the radiosity elements for Gouraud
shading. Shading parameters are passed to the hardware along with the radiosi-
ity elements, with the specular component specified as needed and the diffuse
component turned off. The vertex colors representing the radiosities can typi-
cally be included in the shading equation in the ambient component, by turning
on a white ambient light source. The computed specular shading is then simply
added to the ambient component by the hardware shader. Naturally, highlights
produced by this method will not take shadowing into account. The image of
the boiler room in color plate 40 was produced using this technique.

10.2.8 Surface Reflectance/Transmittance Data

Spectral and bidirectional reflectance and transmittance (BRDF/BTDF) data for
the huge variety of materials found in typical environments is difficult to come
by. Since image synthesis depends on this data, there is growing recognition
of the need to develop fast and inexpensive techniques for making the nec-
essary measurements. The reflectance of a material depends on its chemical
constituents, surface finish, and often subsurface characteristics. In addition,
these characteristics may vary over the surface.
One existing source for spectral data is [202]. This text includes spectral
data for some building materials such as asphalt and brick, and plants such as
lichen. (Also included is the reflectance for assorted foods, including the crust
of baked bread parameterized by thickness!) Some information can also be
found in [237] and [47]. However, these are primarily limited to materials with
important thermal engineering application rather than common architectural
materials. Other sources of material data ranging from metals to sea water to
ripe peaches are [104] and [219]. Surface finish information and related data
on the spatial distribution of reflectance for a few materials can be found in
[104, 114].

Radiosity and Realistic Image Synthesis 324

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.3 PARTICIPATING MEDIA

Drawing on data from [2], Rushmeier [199] provides a sample of reason-

able room values for,

total diffuse reflectances (i.e., averaged over the visible spectrum)

ceiling: 0.60 to 0.90
walls: 0.50 to 0.80
floor: 0.15 to 0.35
furniture: 0.30 (dark wood) to 0.50 (blond wood)

specular reflectances
polished mirror: 0.99
polished aluminum: 0.65

transmission coefficients
clear glass: 0.80 to 0.99 (basically specular)
solid opal glass: 0.15 to 0.40 (basically diffuse.)

Research is also being conducted to develop new methods for measuring the
BRDF of surfaces and for generating the BRDF from simulated surface models
[253, 263].

10.3 Participating Media

The radiosity equation, and the rendering equation from which it is derived,
assumes that light interacts only with surfaces, ignoring the absorption or scat-
tering of light by the medium through which it travels. The assumption of a
nonparticipating medium, normally clear air or vacuum, is reasonable for many
applications. However, clouds, humidity, fog, or smoke are often important
factors in the illumination of exterior environments. For interiors, participating
media may be required in order to simulate phenomena of particular interest, or
for aesthetic reasons (as with the beams of light in the cover image).
A medium affects light transport through absorption, scattering and emis-
sion.9 These may decrease or increase the radiance at any point along a path
through the medium. A fraction a of the radiance L will be absorbed per unit
length along the path. Another fraction, s, of L will be scattered out of the
path (out-scattering). The radiance may also be increased when light incident
on the point from other directions is scattered into the path (in-scattering), or if
the medium itself emits light as in a flame. These phenomena are summarized
in Figure 10.22.

9
This exposition follows that of Rushmeier in [198].

Radiosity and Realistic Image Synthesis 325

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.3 PARTICIPATING MEDIA

absorbtion outscattering emission inscattering

Figure 10.22: Phenomena affecting the radiance entering and exiting a differ-
ential volume along a path through a participating medium.

10.3.1 Path Integrals

The effects of these interactions on the radiance along the path are described by
the differential equation
dL
dS
= tL + aLe + sL (10.27)

where a is the absorption coefficient, s is the scattering coefficient, t is the

extinction coefficient (a + s), L is the radiance along the path, Le is the
emitted radiance, and L is the radiance incident on the path.
The last term of this equation, which accounts for in-scattering, requires
integrating incident radiance over the sphere of incoming directions. The scat-
tering phase function, f(), specifies the fraction of the radiance arriving from
an incident direction, , that is scattered into the path. With the inclusion of the
integral over incoming directions, the equation for radiance in a participating
medium becomes

dL L + L + S
t a e 2
0 0 L ( )f ( ) sin d d (10.28)
ds 4

A number of algorithm have been developed to solve this equation numer-

ically, based on various simplifying assumptions. These are well summarized
in [198]. The following section will concentrate on the generalization of the
radiosity equation to handle media that scatter or emit light isotropically, that is,
with equal radiance in all directions. This reduces the phase function to a con-
stant in a similar fashion to the BRDF reducing to a constant for diffuse surface
reflection. This approach was first introduced to image synthesis by Rushmeier
in [200].
The radiance for a given direction leaving a point in space includes not only
incident light scattered into the path but also light that is just passing through
the point with no interaction at all. The complete radiance at a point is thus

Radiosity and Realistic Image Synthesis 326

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.3 PARTICIPATING MEDIA

expensive to approximate accurately, since the light just passing through a point
typically varies extremely rapidly with direction.
This problem can be avoided by formulating the solution to equation 10.28
in a way that separates out the contributions of in-scattering and emission along
the path. This observation is essentially the same as that used to separate surface
reflection into ideal diffuse and mirror specular terms. Solving equation 10.28
for L(s) gives the integral equation

s
L(s) = L(0)(0, s) + (s, s)J(s)t(s)ds (10.29)
0

where J(s') is a function describing the radiance added to the path at each point
s' due to emission and in-scattering, and (s1, s2) is the integrated extinction
coefficient t along the path from s1 to s2 .
S2
S1 t (s)ds
(s1 , s2 ) = e (10.30)

For an isotropic medium with constant t , (s1, s2) reduces to

(s1 , s2 ) = e t s1 s2 (10.31)

The first term in equation 10.29 accounts for what is seen through the
medium; it consists of the radiance of the surface at the beginning of the path
attenuated by absorption and scattering along the path. This requires only an
approximation of the radiosity at the surface at which the path begins, and thus
eliminates the need to approximate that highly directional component at every
point in the volume.
The second term in equation 10.29 accounts for energy accrued along the
path as the result of in-scattering and emission, which are represented by the
function J(s). J(s), called the source function, varies much more slowly over
the volume than L(s) and can be approximated relatively inexpensively. The
problem is thus to compute J(s).

10.3.2 The Zonal Method

Rushmeiers method of solving for J(s) is based on the zonal method, which
has its origins in the field of heat transfer [130]. The zonal method follows the
development of the standard diffuse surface radiosity closely. J(s) is approxi-
mated by subdividing the volume containing the medium into discrete volume
elements across which the radiance due to scattering or emission is assumed to
be constant. The surfaces are subdivided into surface elements as before (see
Figure 10.23).

Radiosity and Realistic Image Synthesis 327

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.3 PARTICIPATING MEDIA

Figure 10.23: Intersection of volume and surface elements.

Assuming the participating media has been divided into m volume elements
and the surfaces into n elements, then the radiosity of a surface element, i,
includes contributions from all other surfaces, j, and all volume elements, k:
n m

Bi Ai = Ei Ai + i B j F jiSS +
Bk FkiVS (10.32)
j=1 l=1
where F jiSS
is the surface-to-surface form factor and is the volume-to- FkiVS
surface form factor. Likewise, the radiosity of a volume element Vk , includes
contributions from all surface elements j and all other volumes l
n m

4 t Bk V k = 4 a Ek V k + k

j=1
B j F jkSV +
l=1

Bl FlkVV

(10.33)

where F jkSV is the surface-to-volume form factor and FlkVV is the volume-to-
volume form factor. The factor k is the scattering albedo of volume Vk ,
similar to the diffuse reflectivity term for surfaces. The computation of volume-
to-volume and volume-to-surface form factors using the hemicube is shown in
Figure 10.24.
The various form factors are similar to those for conventional radiosity,
except that they include the effect of attenuation due to transport through the

Radiosity and Realistic Image Synthesis 328

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.3 PARTICIPATING MEDIA

Figure 10.24: Computation of volume-to-volume and volume-to-surface from

factors using the hemicube.

medium. The form factor FijSS between two surface elements, Si and Sj , is
given by
cos i cos j
FijSS =

Ai Aj
(x i , x j )
r2
dA j dAi (10.34)

where () (see equation 10.30) reduces the transfer between the respective points
on the two elements i and j. Note also, there is no division by area in this form.
The volume-to-surface form factor is
cos j t,k
FkjVS =
(x , x )
V k Aj
k j
r2
dA j dV k (10.35)

where t,k is the constant t for volume k. The volume-to-volume form factor
is given by
t,k t,l
FklVV =

V k Vl
(x k , x l )
r2
dV l dV k (10.36)

The form factors may be computed using existing algorithms with little
modification. The principle requirement is the evaluation of the path integrals.
In Rushmeiers modified hemicube algorithm, is computed at each hemicube
pixel, based on the distance to the surface or volume to which the form factor
is being computed (see equations 10.30 and 10.31).
Rendering the solution of the zonal method is equivalent to the volume ren-
dering methods developed for scientific visualization applications [75, 136, 149,

Radiosity and Realistic Image Synthesis 329

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 10. EXTENSIONS
10.3 PARTICIPATING MEDIA

150, 264]. Volume rendering for hierarchical volumes is explored in [145]. Dur-
ing rendering, the equation 10.29 is evaluated for each pixel. First, the contribution
of the radiosity of the visible surface is computed, taking into account attenu-
ation by computing as above. Then the contribution to the total radiance of
every volume element through which the path travels is computed, also taking
into account the attenuation based on the distance from the eye to the volume
element. The final radiance for the pixel is the sum of all these contributions.
Since J is computed independently of the view, rendering another view requires
only recomputing these path integrals.
As in conventional radiosity, constant volume elements are not adequate for
rendering. Trilinear interpolation can be used to smooth the volume radiosities
during rendering. Results of this method are shown in color plates 1f and 48.
The zonal method has been extended by Rushmeier [198] to this media
exhibiting weak anisotropic scatter, where volume-to-volume interactions are
assumed insignificant. It has been further extended by Bhate and Tokuta [28]
to the case of more general anisotropic scatter, using spherical harmonics to
approximate the directionally dependent phase function.

Radiosity and Realistic Image Synthesis 330

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH

Chapter 11

Applications and Research

This chapter will explore the use of radiosity in design, engineering and scientific
applications; the experimental validation of the radiosity method; and opportu-
nities for research into improved models, numerical methods, and algorithms.
Applications and research are not as independent as they might seem at first
glance. The incorporation of radiosity into an application precipitates a whole
new set of requirements that can push research in unexpected and fruitful direc-
tions. The difference between algorithms suitable for specialists and algorithms
for mainstream use is not trivial. Since image synthesis is often just a tool
hidden inside a larger application, it is expected to perform robustly and pre-
dictably with little technical input from the user. This places great demands on
the underlying algorithms. The development of automatic, accurate, and robust
algorithms will provide research problems for a long time to come.

11.1 Applications
In general, radiosity is most useful in applications where ideal diffuse interreflec-
tion is important and the geometry is static. The view-independent approach is
particularly valuable when the exploration of a three-dimensional model is re-
quired, rather than the production of a single static image.
When incorporating any global illumination algorithm into an application,
it is important to insulate the user from the technical details of the procedure.
Specifying a mesh resolution or adaptive subdivision criterion forces the user to
learn entirely new concepts unrelated to his or her work. To the extent that the
user is given control over the solution, results must be predictable. Radiosity
solutions for complex environments are too expensive to allow trial and error as
a reasonable approach.
These requirements have important consequences for the development of
radiosity algorithms. Meshing algorithms must be extremely robust. Where
parameters are necessary, they must map clearly to quality and cost. The time-
quality tradeoff should be predictable, and quality should degrade gracefully

Radiosity and Realistic Image Synthesis 331

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

with decreasing cost. For example, a lower-cost solution might provide less
shadow detail, but should not contain disturbing artifacts.
In the next sections, several radiosity applications will be discussed. This
will highlight some of the strengths of the radiosity approach, as well as some
of the areas that need attention.

11.1.1 Architectural Design

Architectural design is in many ways an ideal application for image synthesis
and the radiosity method in particular. Architects must communicate a design to
clients, who may have difficulty visualizing the final building from abstract plans
and drawings. As a result, architects currently depend on hand-drawn or painted
perspectives, or on expensive physical models to provide a concrete represen-
tation. These media allow the designer to communicate a valuable emotional
or aesthetic impression, but they also have limitations. For example, they can
convey only a limited sensation of interior space, since they do not allow the
client to look around or to explore the design from different viewpoints. Because
they are time consuming to produce, they do not encourage the exploration of
alternative materials, finishes, or lighting. For these reasons, the prospect of
producing realistic three-dimensional renderings quickly and automatically has
made image synthesis attractive.
Radiosity is particularly well suited to architectural design. Many interior
surfaces (e.g., upholstery, textiles, matte paints) are reasonably well approxi-
mated by ideal diffuse reflection, and diffuse interreflection makes an important
contribution to the illumination of interiors. Architectural models are usually
static. The radiosity solution, when rendered on a graphics workstation, allows
interactive walkthroughs of the shaded model, giving the designer and the client
the opportunity to explore the interior space interactively.
Typical features of a radiosity application for architecture would include

Translation from modeler data formats.

Access to material libraries.

Access to lighting libraries.

Positioning of lights.

Assignment of material properties.

Positioning of texture maps.

Radiosity and Realistic Image Synthesis 332

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

Figure 11.1: A view of the Lightscape architectural design application. The

model shown is Frank Lloyd Wrights Unity Temple designed in 1904. Image
courtesy of Stuart Feldman Lightscape Graphics Software.

Control over the process of the solution. Progressive refinement is useful

because it allows the user to evaluate the partial solution without having
to wait for convergence.

Camera control for interactive walkthroughs.

Radiosity simulations also have some limitations for architectural visualiza-

tion. Although ideal diffuse reflection is a reasonable approximation for many
common surfaces, many other common surface materials and finishes cannot be
represented correctly (e.g., metals, polishes, glass). Two-pass methods and other
approaches that incorporate specular reflection (see Chapter 10) are important
in architectural visualization. An additional requirement for image synthesis in
general is the availability of data describing materials, finishes, and lights. Rou-
tine use of image synthesis in design applications will require digital libraries
of such data, preferably provided by the manufacturers, analogous to the large
catalogues ubiquitous in design offices.
The cost and performance of radiosity simulation is another issue. Archi-
tectural models can be large; models in excess of 100,000 polygons are not
uncommon. The O(n2) computation and storage cost makes current radiosity
implementations impractical for problems of this size. The problem is intensified

Radiosity and Realistic Image Synthesis 333

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

Figure 11.2: An interior design study. Image courtesy of David L. Munson

Hellmuth Obata & Kassabaum Architects Inc.

by the lack of computing power typically available to small architectural firms.

The time it takes to produce an acceptable image is particularly crucial
because architectural presentation is often the first, rather than the last, stage of
the design process. Presentations of a design proposal to a potential client are
thus developed under pressing time constraints. For image synthesis to play a
role, it must be fast and dependable. Radiosity methods are only beginning to
provide this level of performance.
In spite of these issues, the compelling quality of radiosity images, the
possibility of interactive walkthroughs, and the fact that it is physically based
make radiosity an attractive alternative. An architectural design application
using radiosity is shown in Figure 11.1 and in color plates 47 and 50. The
images in Figures 11.2 and 11.3 were produced using software developed at the
architectural firm of Hellmuth, Obata & Kassabaum, Inc.

11.1.2 Lighting Design

The accurate simulation of global illumination is also a useful tool for lighting
designers. For conventional lighting situations, such as an office, designers often
rely on tables, simple formulae or rules of thumb in deciding how many lighting
fixtures to use and where to position them. For unique lighting designs, a three-
dimensional model may sometimes be constructed and lit. Specialized software

Radiosity and Realistic Image Synthesis 334

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

Figure 11.3: A second view from the same interior design study. Image courtesy
of David L. Munson Hellmuth Obata & Kassabaum Architects Inc.

is also becoming increasingly common. (Wards experimental Radiance package

is a particularly sophisticated general purpose example [252].) Physically based
global illumination models and algorithms provide the possibility of more general
quantitative results as well as realistic images for aesthetic evaluation.
Lighting manufacturers have also begun to develop their own radiosity based
lighting design tools, often with the goal of distributing software along with digi-
tal catalogues of luminaires to architectural and lighting design firms. Companies
ranging from Toshiba Lighting and Technology in Japan, to Philips Electronics
and Zumtobel Licht GmbH in Europe (color plate 51) have begun to experiment
with this technology.

Theatrical Lighting
Theatrical lighting is a special case of lighting design that has very particular
requirements. Although light sources are limited to five or so basic types, a
theatrical production typically uses hundreds of individual lights. These are
turned on and off frequently in different combinations. The continuous variation
of light intensity over time is often aesthetically important.
The circumstances under which the lighting for a particular production is
designed are usually demanding. Sets are constructed at the last minute. As a
result, the actual appearance of the lighted set cannot be fully evaluated until just

Radiosity and Realistic Image Synthesis 335

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

Figure 11.4: Part of a lighting specification for a production at the Metropolitan

Opera House. Each light or group of lights has its own time history of intensity.
Provided by Julie 0B. Dorsey courtesy of the Metropolitan Opera.

prior to the first performance. Computer simulation thus seems ideally suited to
the design of stage lighting.
Several features of the theatrical lighting design problem make radiosity a
potentially useful tool. The model to be illuminated is normally static. The
scene may need to be evaluated from a number of viewpoints throughout the
audience seating. These views are taken from outside the scene, so there is a
fairly well-defined limit to the shading detail required.
On the other hand, there are several aspects of the problem that require
modifications to the conventional radiosity approach. For example, the light
sources have very specific characteristics that must be modeled correctly. The
sources are typically highly directional and thus quite different from ideal diffuse
emitters handled by conventional radiosity. Dorsey [74fldescribes modified form
factor formulae (similar to those discussed in Chapter 10).
Although the geometry of the scene is static, this is not true of the lighting.
The complexity of the problem is suggested by the diagram in Figure 11.4,
which specifies the lighting changes for different banks of lights over time for a
production at the Metropolitan Opera House. Fortunately, most lighting changes
involve changes in emitted power rather than repositioning or repointing of the
light. The static positioning of lights allows the effect of various lights or
combinations of lights to be computed independently. For any given lighting
specification, the independent solutions can then be rapidly scaled appropriately
and summed to provide the total radiosity [73fl.
Dorseys program includes extensive tools for lighting specification (see

Radiosity and Realistic Image Synthesis 336

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

Figure 11.5: The specification and pointing of lights in an application for op-
eratic lighting design. Courtesy of Julie 0B. Dorsey Program of Computer
Graphics Cornell University.

Figure 11.5) and for viewing the solution. Color plates 53 and 54 and the image
on the back cover show results generated by this application. The solutions were
computed using progressive radiosity with ray traced form factors. Rendering
was performed using a stochastic ray tracer. Texture mapping (discussed in
Chapter 9) has also been used very effectively to add detail during rendering.

Lighting Optimization
Kawai et al. [137] have developed techniques that invert the radiosity simulation
paradigm. Given a geometric description of an environment and the illumination
requirements or a desired appearance for the scene, their algorithm addresses
the question of how the lighting parameters should be set to achieve these goals.
They describe a method for designing the illumination in an environment by
applying optimization techniques to a radiosity-based image synthesis system.
The user is able to specify the illumination in terms of subjective criteria such as
pleasantness and privateness. Other goals such as minimizing total energy
consumption or maintaining a minimum illumination on a work surface can
also be specified. An optimization of lighting parameters is then performed

Radiosity and Realistic Image Synthesis 337

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

based on the user-specified constraints and objectives for the illumination of the
environment. The system solves for the best possible settings for light-source
emissivities, element reflectivities, and spot-light directionality parameters.

11.1.3 Remote Sensing

Satellite images of visible and near-infrared light are frequently used to survey
land use and resources. From the point of view of a satellite, a forest or other
region of vegetation is a surface with a characteristic BRDF that determines
its appearance under various observation conditions. One way of determining
this BRDF is to perform a series of measurements in the field.
Computer simulation provides an alternative means of determining the BRDF
For example, to compute a BRDF for a forest canopy, a simplified model of the
canopy specifying its interaction with light can be constructed and evaluated.
One such model treats the canopy as a system of homogeneous volume elements
with a certain density. A model of radiative transfer is then used to compute
the interaction of light with the volumes [99].
Radiosity methods have also been applied to this problem [34, 94]. Since
radiosity requires explicitly modeled geometry, it can support a detailed canopy
model constructed of individual leaves. It thus provides more control over the
characteristics of the simulated vegetation. The view-independence of the ra-
diosity solution is also an advantage. Although the reflectivity of any individual
leaf is treated as ideal diffuse, the BRDF for the canopy as a whole can be highly
directional, due to varying occlusion and anisotropic leaf orientation. Following
the radiosity solution, the BRDF can be evaluated for a range of view angles
by rendering an image of the solution for each view angle. The radiance is
averaged over the pixels of the image, and the BRDF for that view angle is then
determined by the ratio of the radiance to the incident irradiance.
Borel et al. [34] suggest other applications for radiosity in remote sensing,
including modeling the scattering of light between topographic features like the
sides of valleys and the effect of clouds on illumination of land features. They
also envision the use of radiosity in plant biology to simulate light transport and
photosynthesis in plants.

11.1.4 Visual Shape Understanding

The need for a better understanding of visual perception in developing more effi-
cient and accurate image synthesis algorithms has surfaced a number of times in
this book. Interestingly, image synthesis itself can play an important part in im-
proving the understanding of visual perception. For example, texture mapping
and ray tracing have been applied in research into visual shape understand-
ing in experiments to explore how the eye and brain extract information about

Radiosity and Realistic Image Synthesis 338

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.1 APPLICATIONS

geometric shape from the visual field.

Information about geometric shape is provided by binocular vision, as well
as by cues such as texture, shading, cast and received shadows, and highlights
[29, 96, 128, 129, 274]. Wanger [250] has used texture mapping and cast
shadows to explore the importance of these and other cues to positioning tasks in
computer graphics applications. Synthesized images are valuable in investigating
shape and positioning cues because they allow controlled experiments in which
image features are isolated. For similar reasons, image synthesis is useful in
testing algorithms and devices for machine vision. The machine vision field has
relied mainly on edge finding algorithms for object recognition, but researchers
have also investigated shape from shading techniques [128].
Although there is some understanding of how direct illumination provides
shape information, illumination due to interreflected light complicates the prob-
lem [179]. Forsyth and Zisserman [85] have explored the applicability of the
radiosity equation as a model of global illumination for shape perception. They
argue that discontinuities in the radiance function, caused by shadows or surface
creases, provide stronger shape cues than the smooth variation of the function
due to surface curvature. The availability of discontinuity meshing may pro-
vide a means to investigate this possibility. Parker et al. [178] have developed
radiosity code specifically for experimentation in shape understanding.

11.1.5 Infrared Signature Analysis

The goal of infrared signature analysis is the recognition of vehicles or other
objects by their characteristic appearance or signature in images produced by
infrared imaging systems. Computer simulation provides a valuable way to
generate these signatures. An example of a simulated infrared image is shown
in Figure 11.6.
An infrared sensor detects radiance leaving an object due to emission and re-
flection. Infrared imaging requires a complete heat transfer model incorporating
the processes of convection, conduction, and radiative transfer (with shadowing)
for ideal diffuse and specular BRDFs. Heat sources include engines, exhaust,
solar illumination, and sky illumination.
Many of these requirements are outside of the concerns of normal image syn-
thesis. However, the radiative transfer component can benefit from algorithms
developed for radiosity as applied to image synthesis, particularly algorithms for
computing form factors. For example, the hemicube algorithm has been used
by Johnson et al. [133] for this application.

Radiosity and Realistic Image Synthesis 339

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.2 EXPERIMENTAL VALIDATION

Figure 11.6: A simulated infrared image. Courtesy of Keith R. Johnson, Michi-

gan Technological University.

11.1. 6 Fine Arts

Computer graphics has already made an impact on the fine arts. Conferences
such as SIGGRAPH include computer art and film shows in addition to technical
presentations. Although much of this work is abstract, the ability to simulate
realistic lighting effects can be an effective tool for conveying meaning in artistic
works. Color plate 52 depicts a frame from an experimental video by Monika
Fleischmann entitled Home of the Brain, in which lighting effects are simu-
lated with radiosity.

11.2 Experimental Validation

As applications turn increasingly to global illumination algorithms for design,
engineering, or scientific problems, there is a corresponding need to verify that
the algorithms produce valid results. Radiosity is a physically based model and
purports to produce a physically accurate description of illumination. This is a
claim that can be tested by comparing the results of a radiosity simulation with
measurements of real models.
Experimental validation addresses two basic questions. First, does the math-
ematical model adequately describe the physical process of interest? For exam-

Radiosity and Realistic Image Synthesis 340

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.2 EXPERIMENTAL VALIDATION

ple, in the case of radiosity, one might ask whether the assumption of ideal
diffuse reflection is too restrictive to be useful for modeling real environments.
Second, experimental validation tests the accuracy of numerical methods and
algorithms.
A further issue that requires experimental investigation is the quality of
the visual experience provided by image synthesis algorithms. The perceived
realism of images is very difficult to quantify, but this is nevertheless a crucial
criterion in the evaluation of algorithms for certain applications.
Meyer et al. [167] have made preliminary efforts toward an experimental
validation of the radiosity method. They have performed two sets of experiments
with the familiar red, white, and blue box (similar to the boxes in color plate 1).
The first set of these experiments compares the results of a radiosity simula-
tion to radiometric measurements of a physical model. These experiments were
intended to evaluate basic assumptions of the radiosity method, particularly the
assumption of ideal diffuse reflection and emission. Measurements of the inte-
gral of irradiance were made at intervals on the open face of a box whose interior
faces were painted various colors. Illumination was provided by a diffused light
source at the center of the top of the box.
The results of the experiment indicate that the assumption of ideal diffuse
emission for the light source contributes significantly to errors in the simulation.
The experiment also demonstrates that radiosity accounts well for the effect of
diffuse interreflection or color bleeding and for the effect of occlusion, insofar
as measurement by integrating irradiance can indicate.
The second set of experiments addresses the perceived quality of radiosity
images, with the intention of seeing how close radiosity comes to creating a
visual experience identical to that experienced seeing the real .scene. Given the
limitations of existing display devices, it was necessary to simplify the problem
in several ways. For example, the light source was blocked from direct view
in both the image and the actual scene so that the limited dynamic range of the
display did not immediately give away the simulation.
In the experiment, a real scene and a simulated version were projected side
by side onto frosted glass in the backs of two view cameras (see Figure 11.7 and
color plate 5). Subjects were asked to determine which was the real scene and
to rate the relative quality of features like shadows and color (color plates 6 and
7). The simulated image was chosen as the real scene approximately 50% of the
time (for untrained observers), indicating that the synthesized image achieved
the goal of perceptual realism in this highly constrained case. Color plates 8
and 9 are the result of exposing film placed in the view cameras. It should be
noted that the reader is not getting the same visual experience as the subjects
in the experiment. Film processing introduces its own distortions, as does the
reproduction of the photograph in print. In addition, viewing light reflected from

Radiosity and Realistic Image Synthesis 341

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.2 EXPERIMENTAL VALIDATION

Figure 11.7: The experimental setup for perceptual validation of a simple ra-
diosity solution.

the pages of a book is different than viewing light transmitted through frosted
glass. Finally, the images presented to the experimental observer were upside
down due to camera lens optics.
These experiments are clearly just a beginning. However, they raise many of
the issues that any experimental validation must address and provide preliminary
answers to some of them. These issues include the type of radiometric device
appropriate for the physical measurements, the physical quantities that should be
measured, and the problem of constructing a physical model and measuring its
properties to define an equivalent computer model. For perceptual experiments,
there are additional difficult questions, such as how to present the comparisons
without biasing the result and how to obtain detailed information about what
features contribute or detract from realism.

Radiosity and Realistic Image Synthesis 342

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.3 FUTURE RESEARCH DIRECTIONS

Ward [251] has also performed experimental verification of the Radiance

lighting design program. Although this program does not use the radiosity
method, the issues of experimental verification are essentially the same and a
number of practical questions are addressed in this work. In spite of these
contributions, experimental validation is clearly an area where much valuable
work remains to be done.

11.3 Future Research Directions

Although developments over the last several years have demonstrated that radios-
ity is potentially a useful tool for image synthesis, fundamental improvements
still remain to be made to ensure accuracy, achieve predictable results, and pro-
vide adequate performance. Several key areas for continuing work are described
briefly in the following sections.

11.3.1 Error Analysis

The numerical evaluation of a mathematical model will rarely, if ever, produce
an exact result. This is not necessarily a handicap, as long as the error can
be characterized quantitatively and can be reduced to an acceptable level given
enough computational resources. For radiosity applications, it would be ideal to
bound the maximum error for the radiosity values in an image to, for example, to
within 5%. In reality, this is very difficult to achieve, because error is introduced
at many different points during the solution and propagates in a complicated
manner.
Error is introduced by interpolation, the approximations made in discretiza-
tion, numerical integration of form factors, the inability to converge the matrix
solution fully, and the projection of the solution onto a new basis prior to render-
ing. Future work must develop a better understanding of where and how error is
introduced into the radiosity solution, as well as how these errors accumulate to
cause inaccuracies in the solution. This research will undoubtedly benefit from
the existing literature in numerical methods and engineering analysis. Progress
in placing global illumination analysis on a rigorous mathematical footing is also
a prerequisite to understanding the sources and propagation of errors.

11.3.2 Perceptually Based Error Metrics

As a number of researchers have recognized [167, 238, 250], the image synthesis
process does not end with an approximation of the radiosity or radiance function
for the scene. The evaluation of a global illumination model is only a step in a
process whose final outcome is a visual experience in the mind of the viewer.

Radiosity and Realistic Image Synthesis 343

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.3 FUTURE RESEARCH DIRECTIONS

Thus, although it is important to develop quantitative error measures, a full

evaluation of the simulation must account for the perception of the image. In
this respect, the job of image synthesis is more difficult than that of engineering,
because the problem to be solved includes a perceptual process for which there
are no complete and well-defined mathematical models.
The ability to measure the quality of the perceived image is important for
more than evaluating the relative usefulness of different algorithms. A measure
of accuracy is a key tool for any numerical algorithm that attempts to direct the
course of the solution in order to minimize error and distribute computational
resources efficiently. For example, a perceptually based error metric might di-
rect the radiosity solution to focus effort on aspects of shadow boundaries that
are important to perceived quality, while ignoring others. A purely radiomet-
ric measure might weight these characteristics equally. Currently, assumptions
about which aspects of reality are perceptually important are most often built
into the algorithm or the mathematical model on the basis of what i.s practical
to compute rather than on firm perceptual grounds.
One goal of research in this area is to develop a better understanding of
visual perception, with the ultimate aim of modeling the aspects of the process
that are relevant to image synthesis. There is a large literature of perception
research to be investigated. The literature of the fine arts is also a resource
that should not be overlooked. Artists have struggled with many of these issues
for centuries, and some of the discoveries of computer graphics research are
available in art handbooks on the shelves of the local bookstore.
A related area for research is the development of algorithms that can incor-
porate image-based error measures into the solution process. The importance
algorithm of Smits et al. [220] (Chapter 7) is a valuable first step in this direction.

11.3.3 Physically Based Emission and brdf Data

The results of a physically-based global illumination algorithm are only as accu-
rate as the data describing the model. Radiosity requires a geometric description
of the environment, emission spectra for light sources, directional luminance
data for light fixtures, reflectivity spectra for surfaces and, for more general al-
gorithms, the BRDF for surfaces. The generation of geometric models is a field
in its own right and is not covered in this text. However, the physically accurate
specification of light emission and surface reflectivity is beginning to receive
attention as a topic for research in the global illumination field. A survey of
sources for data on a limited set of materials and lighting is given by Rushmeier
[199].
Faced with the unavailability of the full range of data required for image
synthesis, researchers have begun to investigate techniques for measuring the

Radiosity and Realistic Image Synthesis 344

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.3 FUTURE RESEARCH DIRECTIONS

BRDF of real surfaces and for generating the BRDF from simulated surface
models. Ward [253] discusses an experimental setup for measuring the BRDF
of real materials. Cabral, et al. [41] describe a method for computing the BRDF
from bump maps specifying small scale surface geometry. Westin et al. [263]
describe an algorithm for computing the BRDF based on detailed geometric
models of materials such as cloth or brushed metal. As image synthesis algo-
rithms become more accurate, their dependence on valid data will become more
critical and this area of research will gain increasing importance.

11.3.4 Meshing
As evident from the discussion of applications, there is a great need for robust,
fast, and predictable meshing algorithms. One of the largest hindrances to the
adoption of radiosity for mainstream applications is the difficulty of specifying
meshing parameters (and the trial and error required to get good results).
Meshing for radiosity has been until recently, to use Heckberts words,
something of a black art [120]. Although this state of affairs has improved,
as demonstrated by the survey of meshing algorithms in Chapters 6 and 8, there
are many techniques for finite element meshing that have yet to be investigated
for radiosity.
Discontinuity meshing has made impressive gains in improving predictabil-
ity and in eliminating the need for trial and error in setting mesh parameters.
The development of discontinuity meshing algorithms that are efficient and prac-
tical for complex environments is an important direction to investigate. There
is also room to improve a posteriori approaches, which may reduce the need to
determine exact boundaries a priori.

11.3.5 Hierarchy
As noted in section 11.1.1, architectural models, for which radiosity is otherwise
well suited, can easily exceed 100,000 polygons. Despite many improvements
in efficiency, radiosity algorithms are still essentially O(n2) in the number of
polygons. Fortunately, this does not seem to be an inherent lower bound on the
efficiency of the technique.
The hierarchical methods described in Chapter 7 support the view that the
cost of radiosity can be reduced. The basic concept underlying these methods
is that the interactions between two widely separated groups of elements can
be represented by a single average interaction. Current hierarchical algorithms
always start with a polygon and subdivide downward. Thus elements can be
grouped only within a polygon, and elements from different polygons cannot
be grouped together to form larger clusters. As a result, current hierarchical
methods reduce the cost of the radiosity solution in terms of the number of

Radiosity and Realistic Image Synthesis 345

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.3 FUTURE RESEARCH DIRECTIONS

elements, but the cost in the number of polygons is still O(n2). Clearly, the
next step is to cluster surfaces together, allowing interactions with multiple
objects to be treated as a single interaction where appropriate. Preliminary work
has begun in this area.

Xu [269] has used spatial substructuring to reduce the cost of the radiosity
solution. In this approach, the volume containing the environment is subdi-
vided uniformly to create subscenes. Elements within a single subscene interact
normally, but interactions between elements in neighboring subscenes are medi-
ated by special elements on the surfaces that define the subscene boundaries.
These elements record directional information about the light reflected toward
the boundary from inside the subscene. For the neighboring subscene, these
boundary elements then provide an approximation of the illumination arriving
from outside the subscene to the elements within. This approach has been used
subsequently as a basis for the parallelization of the radiosity solution [10, 239],
with each independent subscene handled by a different processor. A related
approach has also been proposed by Neumann and Kelemen [172].

Although spatial subdivision reduces the cost of the radiosity solution, the
subdivision must be specified a priori by the user. The hierarchy is inflexible and
is limited to two levels, which does not fully exploit the potential of hierarchical
methods to increase the efficiency of the solution.

Rushmeier et al. [196] describe a more flexible algorithm based on a two-

pass method. In the first pass a radiosity solution is performed on a simplified
version of the scene, in which complex objects are approximated by bounding
boxes. The criterion for simplification is based on an approximation to the
solid angle subtended by surfaces, objects, and groups of objects with respect to
potential receiving surfaces. The rendering pass uses Monte Carlo ray tracing
from the eye to compute direct illumination. Secondary rays are shot to compute
the indirect illumination of visible surfaces. Beyond a certain neighborhood
around the surface to be shaded, the illumination is provided by the radiosity
solution computed in the first pass, which effectively clusters geometric detail
into simplified representations. Even though the hierarchy of complex and simple
geometric representation is limited to two levels, Rushmeier reports decreases of
several orders of magnitude in solution time over conventional radiosity methods.

The clustering algorithms developed so far are only a first step. Future work
should include automatic methods for generating the simplified geometry and
generalization of the hierarchy to multiple levels. This remains perhaps the most
important open problem for radiosity research.

Radiosity and Realistic Image Synthesis 346

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.4 CONCLUSION

Figure 11.8: A view looking down the nave of Chartres Cathedral. The model
was produced by John Lin using Hewlett-Packards ME-30 solid modeler.

11.4 Conclusion
Architects have struggled to capture the grandeur of space and light in monu-
mental structures. The renderings of such a structure in Figures 11.8 and 11.9
demonstrate that the radiosity method can produce compelling images based on
the simulation of the propagation of light.
The goal of producing convincingly realistic images of nonexistent scenes is
tantalizingly close to being achieved. However, many chapters in this exciting
research endeavor have yet to be written. As computational speeds increase
and the associated costs continue to decline, emerging applications will surely
demand more of image synthesis algorithms. We hope that this text has provided
a firm base upon which to build these future image synthesis systems and that
it will encourage the reader to become involved in this quest.

Radiosity and Realistic Image Synthesis 347

Edited by Michael F. Cohen and John R. Wallace
CHAPTER 11. APPLICATIONS AND RESEARCH
11.4 CONCLUSION

Figure 11.9: An exterior view of the Chartres model.

Radiosity and Realistic Image Synthesis 348

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

Bibliography
[1] CIE Recommendations on Uniform Color Spaces, Color-difference Equa-
tions, and Psychometric Color Terms. Bureau de la CIE, Paris, 1978.

[2] IES Lighting Handbook. 1981 Reference Edition, New York, 1981.

[3] Sweets Catalog File: Products for Engineering. McGraw-Hill, New York,
1981.

[4] ANSI standard nomenclature and definitions for illuminating engineering.

ANSI/IES RP-16-1986, Illuminating Engineering Society, 345 East 47th
Street, New York, NY 10017, approved 1986.

[5] AIREY , J. M., ROHLF , J. H., and BROOKS, F. P. Towards image

realism with interactive update rates in complex virtual building environ-
ments. Computer Graphics (1990 Symposium on Interactive 3D Graphics)
24:2 (Mar. 1990), pp. 4150.

[6] ALA, S. R. Performance anomalies in boundary data structures. IEEE

Computer Graphics and Applications 12:2 (Mar. 1992), pp. 4958.

[7] AMANATIDES, J. Ray tracing with cones. Computer Graphics (SIG-

GRAPH 84 Proceedings) 18:3 (July 1984), pp. 129135.

[8] APPEL, A. Some techniques for shading machine renderings of solids. In

Proceedings of the Spring Joint Computer Conference (1968), pp. 3745.

[9] APPEL, A. A. An efficient program for many body simulation. Siam

Journal of Sci. Stat. Computing 6:1 (1985), pp. 85103.

[10] ARNALDI , B., PUEYO, X., AND VILAPLANA, J. On the division

of environments by virtual walls for radiosity computation. In Second
Eurographics Workshop on Rendering (Barcelona, Spain, May 1991).

[11] ARVO, J., and KIRK, D. Fast ray tracing by ray classification. Com-
puter Graphics (SIGGRAPH 87 Proceedings) 21:4 (Aug. 1987), pp. 55
64.

Radiosity and Realistic Image Synthesis 349

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[12] ARVO, J. R. Backward ray tracing. In Developments in Ray Tracing,

SIGGRAPH 86 course notes, Vol. 12. Aug. 1986.

[13] ATHERTON, P. R., WEILER, K., AND GREENEBERG, D. Polygon

shadow generation. Computer Graphics (SIGGRAPH 78 Proceedings)
12:3 (l978), pp. 275281.

[14] AUPPERLE , L., AND HANRAHAN, P. A hierarchical illumination algo-

rithm for surfaces with glossy reflection. Computer Graphics (SIGGRAPH
93 Proceedings) 27 (Aug. 1993).

[15] BABUSKA, I., ZIENKIEWICZ, 0., GAGO, J., AND DE A. OLIVEIRA,

E., Eds. Accuracy Estimates and Adaptive Refinements in Finite Element
Computations. John Wiley and Sons, New York, 1986.

[16] BARNES, J., AND HUT, P. A hierarchical O(n log n) force calculation
algorithm. Nature 324 (1986), pp. 446449.

[17] BARTELS , R. H., BEATTY, J. C., AND B ARSKY, B. A. An Introduc-

tion to Splines for Use in Computer Graphics and Geometric Modeling.
Morgan Kaufmann, Los Altos, Calif., 1987.

[18] BAUM, D. R., MANN, S., SMITH, K. P., AND WINGET , J. M. Mak-
ing radiosity usable: Automatic preprocessing and meshing techniques for
the generation of accurate radiosity solutions. Computer Graphics (SIG-
GRAPH 91 Proceedings) 25:4 (July 1991), pp. 5160.

[19] BAUM, D. R., RUSHMEIER , H. E., AND WINGET, J. M. Improv-

ing radiosity solutions through the use of analytically determined form-
factors. Computer Graphics (SIGGRAPH 89 Proceedings) 23:3 (July
1989), pp. 325334.

[20] BAUM, D. R., WALLACE, J. R., C OHEN, M. F., AND GREENBERGE,

D. P. The back-buffer algorithm: an extension of the radiosity method to
dynamic environments. The Visual Computer 2:5 (Sept. 1986), pp. 298
306.

[21] BAUM, D. R., AND WINGET, J. M. Real time radiosity through

parallel processing and hardware acceleration. Computer Graphics (1990
Symposium on Interactive 3D Graphics) 24:2 (Mar. 1990), pp. 6775.

[22] BAUMGART, B. G. A polyhedron representation for computer vision.

In AFIPS Conference Proceedings (1975), Vol. 44, pp. 589596.

Radiosity and Realistic Image Synthesis 350

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[23] BECKER, E. B., CAREY, G. F., AND ODEN, J. T. Finite Elements,

An Introduction, Volume 1. Prentice Hall, Englewood Cliffs, NJ, 1981.

[24] BERAN-KOEHN, J. C., AND PAVICIC, M. J. A cubic tetrahedral

adaptation of the hemi-cube algorithm. In Graphics Gems II, J. Arvo, Ed.
AP Professional, Boston, San Diego, 1991, pp. 299302.

[25] BERAN-KOEHN, J. C., AND PAVICIC, M. J. Delta form-factor calcu-

lation for the cubic tetrahedral algorithm. In Graphics Gems III, D. Kirk,
Ed. AP Professional, Boston, 1992, pp. 32328.

[26] BEYLKIN, G., COIFMAN, R., AND ROKHLIN, V. Fast wavelet trans-
forms and numerical algorithms I. Communications on Pure and Applied
Mathematics 44 (1991), pp. 141183.

[27] BEYLKIN, G., COIFMAN, R., AND ROKHLIN, V. Wavelets in numeri-

cal analysis. In Wavelets and Their Applications, G. Beylkin, R. Coifman,
I. Daubechies, S. Mallat, Y. Meyer, L. Raphael, and B. Ruskai, Eds. Jones
and Bartlett, Cambridge, 1992, pp. 181210.

[28] BHATE, N., AND TOKUTA, A. Photorealistic volume rendering of

media with directional scattering. In Third Eurographics Workshop on
Rendering (Bristol, UK, May 1992), pp. 227245.

[29] BLAKE, A., AND BRELSTAFF , G. Geometry from specularity. In

Proceedings of the Second International Conference on Computer Vision
(Tampa Springs, FL, 1988), pp. 394403.

[30] BLINN, J. F. Models of light reflection for computer synthesized pic-

tures. Computer Graphics (SIGGRAPH 77 Proceedings) 11:2 (1977),
pp. 192198.

[31] BLINN, J. F. Simulation of wrinkled surfaces. Computer Graphics

(SIGGRAPH 78 Proceedings) 12:3 (Aug. 1978), pp. 286292.

[32] BLINN, J. F., AND NEWELL, M. E. Texture and reflection in computer

generated images. Communications of the ACM 19:10 (1976), pp. 542
547.

[33] BOENDER, E. Finite Element Mesh Generation from CSG Models. PhD
thesis, Dept. of Technical Math. and Informatics, Delft University of Tech-
nology, Netherlands, 1992.

Radiosity and Realistic Image Synthesis 351

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[34] BOREL, C. C., GERSTL, S. A. W., AND POWERS, B. J. The radiosity

method in optical remote sensing of structured 3-d surfaces. Remote Sens.
Environ. 36 (1991), pp. 13-44.

[35] BOUGUER, P. The Gradation of Light. University of Toronto Press,

1960.

[36] BOUKNIGHT, J., AND KELLEY, K. An algorithm for producing half-

tone computer graphics presentations with shadows and movable light
sources. In Proceedings of the Spring Joint Computer Conference, AFIPS
(1970), Vol. 36, AFPIS Press, pp. 110.

[37] BOUKNIGHT, W. J. A procedure for generation of three-dimensional

half-toned computer graphics presentations. Communications of the ACM
13:9 (Sept. 1970), pp. 292301.

[38] BREBBIA. C. A., AND DOMINGUEZ J. Boundary Elements: An Intro-

ductory Course. McGraw-Hill, New York, 1992.

[39] BU, J., AND DEPRETTERE . E. F. A VLSI system architecture for

high-speed radiative transfer 3D image synthesis. The Visual Computer|
5:3 (June 1989), pp. 121133.

[40] BUCKALEW, C., AND F USSELL, D. Illumination networks: Fast re-

alistic rendering with general reflectance functions. Computer Graphics
(SIGGRAPH 89 Proceedings) 23:3 (July 1989), pp. 8998.

[41] CABRAL, B., MAX, N., AND SPRINGMEYER, R. Bidirectional reflec-

tion functions from surface bump maps. Computer Graphics (SIGGRAPH
87 Proceedings) 21:4 (July 1987), pp. 273281.

[42] CAMPBELL, A. Modeling Global Diffuse Illumination for Image Syn-

thesis. PhD thesis, Dept. of Computer Sciences, University of Texas at
Austin, Dec. 1991.

[43] CAMPBELL, A., AND FUSSELL, D. S. Adaptive mesh generation for

global diffuse illumination. Computer Graphics (SIGGRAPH 90 Pro-
ceedings) 24:4 (Aug. 1990), pp. 155164.

[44] CATMULL, E. Computer display of curved surfaces. In Proceedings

of the IEEE Conference on Computer Graphics, Pattern Recognition, and
Data Structures (May 1975), Vol. 11, pp. 1117.

Radiosity and Realistic Image Synthesis 352

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[45] CENDES, Z. J., AND WONG, S. H. C 1 quadratic interpolation over

arbitrary point sets. IEEE Computer Graphics and Applications 7:11 (Nov.
1987), pp. 816.

[46] CHALMERS, A. G., AND PADDON, D. J. Parallel processing of pro-

gressive refinement radiosity methods. In Second Eurographics Workshop
on Rendering (Barcelona, Spain, May 1991).

[47] CHANEY, J., RAMIDAS, V., RODRIGUEZ, C., AND WU, M., Eds.
Thermophysical Properties Research Literature Retrieval Guide 1900-
1980. IFI/Plenum, New York, 1982.

[48] CHATTOPADHYAY, S., AND FUJIMOTO, A. Bi-directional ray tracing.

In Computer Graphics 1987 (Proceedings of Computer Graphics Interna-
tional 87) (Tokyo, 1987), Springer-Verlag, pp. 335343.

[49] CHEN, H., AND WU, E.-H. An efficient radiosity solution for bump
texture generation. Computer Graphics (SIGGRAPH 90 Proceedings)
24:4 (Aug. 1990), pp. 125134.

[50] CHEN, S. E. A progressive radiosity method and its implementation in a

distributed processing environment. Masters thesis, Program of Computer
Graphics, Cornell University, Jan. 1989.

[51] CHEN, S. E. Incremental radiosity: An extension of progressive radiosity

to an interactive image synthesis system. Computer Graphics (SIGGRAPH
90 Proceedings) 24:4 (Aug. 1990), pp. 135144.

[52] CHEN, S. E., RUSHMEIER, H. E., MILLER , G., AND TURNER, D. A

progressive multi-pass method for global illumination. Computer Graph-
ics (SIGGRAPH 91 Proceedings) 25:4 (July 1991), pp. 164174.

[53] CHEW, L. Constrained Delaunay triangulations. In 3rd Symp. Comp.

Geom. (1987), pp. 215222.

[54] CHIN, N., AND FEINER, S. Near real-time shadow generation using
bsp trees. Computer Graphics (SIGGRAPH 89 Proceedings) 23:3 (July
1989), pp. 99106.

[55] CHIN, N., AND FEINER, S. Fast object-precision shadow generation for
area light source using bsp trees. In Computer Graphics, Special Issue
(Proceedings 1992 Symposium on Interactive 3D Graphics) (Cambridge,
Mass., Mar. 1992), ACM Press, pp. 2130.

Radiosity and Realistic Image Synthesis 353

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[56] CLOUGH, R., AND TOCHER, J. Finite element stiffness matrices for
analysis of plate bending. In Matrix Methods in Structural Mechanics
(Proceedings of the conference held at Wright-Patterson Air Force Base,
Ohio, 26-28 October 1965) (1966), pp. 515545.

[57] CLOUGH, R. W. The finite element in plane stress analysis. In Procee-

dings of the Second ASCE Conference on Electronic Computation (Sept.
1960).

[58] COHEN, M. A radiosity method for the realistic image synthesis of

complex diffuse environments. Masters thesis, Program of Computer
Graphics, Cornell University, Aug. 1985.

[59] COHEN, M., CHEN, S. E., W ALLACE, J. R., AND GREENBERG,

D. P. A progressive refinement approach to fast radiosity image gen-
eration. Computer Graphics (SIGGRAPH 88 Proceedings) 22:4 (Aug.
1988), pp. 7584.

[60] COHEN, M., AND G REENBERG, D. P. The hemi-cube: A radiosity

solution for complex environments. Computer Graphics (SIGGRAPH 85
Proceedings) 19:3 (Aug. 1985), pp. 3140.

[61] COHEN, M., GREENBERG, D. P., IMMEL, D. S., AND BROCK,

P. J. An efficient radiosity approach for realistic image synthesis. IEEE
Computer Graphics and Applications 6:3 (Mar. 1986), pp. 2635.

[62| COHEN, M. F., AND GREENBERG, D. P. The hemi-cube: A radiosity

solution for complex environments. Computer Graphics (SIGGRAPH 85
Proceedings) 19:3 (July 1985), pp. 3140.

[63] COOK, R. L. Stochastic sampling in computer graphics. ACM Transac-

tions on Graphics 5:1 (Jan. 1986), pp. 5172.

[64] COOK, R. L., PORTER, T., AND CARPENTER , L. Distributed ray

tracing. Computer Graphics (SIGGRAPH 84 Proceedings) 18:3 (July
1984), pp. l37145.

[65] COOK, R. L., AND TORRANCE, K. E. A reflection model for computer

graphics. ACM Transactions on Graphics 1:1 (1982), pp. 724.

[66] COURANT, R. Variational methods for the solution of problems of equi-

librium and vibration. Bulletin of the American Mathematical Society 49
(1943), pp. 123.

Radiosity and Realistic Image Synthesis 354

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[67] CROW, F. C. Shadow algorithms for computer graphics. Computer

Graphics (SIGGRAPH 77 Proceedings) 11:2 (1977), pp. 242248.

[68] DAVISON, B. Neutron Transport Theory. Oxford University Press, 1957.

[69] DELVES, L. M., AND MOHAMED, J. L. Computational methods for

integral equations. Cambridge University Press, Cambridge, UK, 1985.

[70] DESAI , C. S., AND ABEL, J. F. Introduction to the Finite Element

Method. Van Nostrand Reinhold, New York, 1972.

[71] DIPPE, M. A. Z., AND WOLD, E. H. Stochastic sampling: Theory

and application. In Progress in Computer Graphics, G. W. Zobrist, Ed.
Ablex Publishing, Norwood, NJ, 1991.

[72] DOCTOR, L., AND TORBORG, J. Display techniques for octree-encoded

objects. Ieee Computer Graphics and Applications 1:3 (July 1981),
pp 2938.

[73] DORSEY, J. O. Computer Graphics Techniques for Opera Lighting De-

sign and simulation. PhD thesis, Program of Computer Graphics, Cornell
University, Jan. 1993.

[74] DORSEY, J. O., SILLION, F. X., AND G REENBERG, D. P. Design and

simulation of opera lighting and projection effects. Computer Graphics
(SIGGRAPH 91 Proceedings) 25:4 (July 1991), pp. 4150.

[75] DREBIN , R. A., CARPENTER, L., AND HANRAHAN, P. Volume

rendering. Computer Graphics (Proceedings of SIGGRAPH 1988) 22:4
(August 1988), pp. 6574.

[76] DRETTAKIS, G., AND FIUME, E. Structure-directed sampling, recon-

struction, and data representation for global illumination. Second Euro-
graphics Workshop on Rendering (May 1991), pp. 189201.

[77] DRUCKER, S. M., AND SCHRDER, P. Fast radiosity using a data par-
allel architecture. In Third Eurographics Workshop on Rendering (Bristol,
uk, May 1992), pp. 247258.

[78] EMERY, A. F., JOHANSSON, O., LOBO, M., AND ABROUS, A. A

comparative study of methods for computing the diffuse radiation view-
factors for complex structures. The Journal of Heat Transfer 113 (May
1991), pp. 413422.

Radiosity and Realistic Image Synthesis 355

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[79] ESSELINK, E. About the order of Appels algorithm. Computing Science

Note KE5-1, University of Groningen, 1989.

[80] FARIN, G. Curves and Surfaces for Computer Aided Geometric Design.
Academic Press, Boston, 1988.

[81] FEDA, M., AND PURGATHOFER , W. Progressive refinement radiosity

on a transputer network. In Second Eurographics Workshop on Rendering
(Barcelona, Spain, May 1991).

[82] FEDA, M., AND PURGATHOFER, W. Accelerating radiosity by over-

shooting. In Third Eurographics Workshop on Rendering (Bristol, UK,
May 1992), pp. 2132.

[83] FOCK, V. A. Illumination produced by surfaces of arbitrary shape.

Proceedings of the State Optical Institute 3:28 (1924).

[84] FOLEY, J. D., VAN D AM, A., F EINER, S. K., AND HUGHES, J. F.
Computer Graphics, Principles and Practice, 2nd Edition. Addison-
Wesley, Reading, Massachusetts, 1990.

[85] FORSYTH, D., AND ZISSERMAN, A. Shape from shading in the light
of mutual illumination. Image and Vision Computing 8:1 (Feb. 1990),
pp. 4249.

[86] FRANKE, R. Scattered data interpolation: Tests of some methods. Math-

ematics of Computation 38:157 (Jan. 1982), pp. 181200.

[87] FUCHS, H., K EDEM, Z. M., AND NAYLOR, B. F. On visible surface

generation by a priori tree structures. Computer Graphics (SIGGRAPH
80 Proceedings) 14:3 (July 1980), pp. 124133.

[88] FUCHS, H., POULTON, J., EYLES, J., G REER, T., GOLDFEATHER,
J., ELLSWORTH , D., MOLNAR, S., TURK, G., TEBBS, B., AND
ISRAEL, L. Pixel-planes 5: A heterogeneous multiprocessor graphics
system using processor-enhanced memories. Computer Graphics (SIG-
GRAPH 89 Proceedings) 23:3 (July 1989), pp. 7988.

[89] FUNKHOUSER, T. A., SEQUIN, C. H., AND TELLER, S. J. Man-

agement of large amounts of data in interactive building walkthroughs.
In Computer Graphics, Special Issue (Proceedings 1992 Symposium on
Interactive 3D Graphics, Cambridge, Mass. 29 Mar, 1992) (Mar. 1992),
ACM Press, pp. 11-20.

[90] GASTINEL, N. Linear NumericalAnalysis. Academic Press, 1970.

Radiosity and Realistic Image Synthesis 356

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[91] GEORGE, D. W., SILLION, F. X., AND GREENBERG, D. P. Radiosity

redistribution for dynamic environments. IEEE Computer Graphics and
Applications 10:4 (July 1990), pp. 2634.

[92] GEORGE, P. Automatic Mesh Generation. Wiley, New York, 1991.

[93] GERSHUN, A. The Light Field. Moscow, 1936. Translated in Journal of

Mathematics and Physics, Vol. 18, No. 2, 1939.

[94] GERSTL, S. A. W., AND B OREL, C. C. Principles of the radiosity

method for canopy reflectance modeling. In International Geoscience and
Remote Sensing Symposium, 2024 May 1990, Washington, dc, Proceed-
ings IGARSS '90 (May 1990), Vol. 3, pp. 1735-1737.

[95] GIFFORD, S. Data Parallel Two Pass Rendering. Naval Research Lab-
oratory Technical Report, Aug. 1991.

[96] GILSHRIST, A. L., AND JACOBSEN, A. Perception of lightness and

illumination in a world of one reflectance. Perception 13 (1984), pp. 5
19.

[97] GLASSNER, A. S., Ed. An Introduction to Ray Tracing. Academic Press,

San Diego, 1989.

[98] GLASSNER, A. S. Maintaining winged-edge models. In Graphics Gems

III, D. Kirk, Ed. AP Professional, Boston, 1992, pp. 191201.

[99] GOEL, N. S. Models of vegetation canopy reflectance and their use

in estimation of biophysical parameters from reflectance data. Remote
Sensing Review 4:1 (1988), p. 221.

[100] GORAL, C. M., TORRANCE, K. E., GREENBERG, D. P., AND

BATTAILE, B. Modelling the interaction of light between diffuse sur-
faces. Computer Graphics (SIGGRAPH 84 Proceedings) 18:3 (July
1984), pp. 212222.

[101] GORTLER, S., AND COHEN, M. F. Radiosity and relaxation methods.

Technical Report TR 40893, Princeton University, 1993.

[102] GORTLER, S. J., SCHRDER, P., COHEN, M. F., AND HANRAHAN,

P. M. Wavelet radiosity. Computer Graphics (SIGGRAPH 93 Proceed-
ings) 27 (Aug. 1993).

[103] GOURAUD, H. Computer display of curved surfaces. Tech. rep., Dept.

of Computer Science, University of Utah, Salt Lake City, Utah, 1971.

Radiosity and Realistic Image Synthesis 357

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[104] GRAY, D. E., Ed. American Institute of Physics Handbook. McGraw

Hill, New York, 1972.

[105] GREENBERG, D. P., COHEN, M., AND TORRANCE, K. E. Radiosity:

A method for computing global illumination. The Visual Computer 2:5
(Sept. 1986), pp. 29 l297.

[106] GREENGARD, L. The Rapid Evaluation of Potential Fields in Particle

Systems. MIT Press, 1988.

[107] GUITTON, P., ROMAN, J., AND S CHLICK, C. Two parallel approaches
for a progressive radiosity. In Second Eurographics Workshop on Render-
ing (Barcelona, Spain, May 1991).

[108] HAINES, E,. Essential ray tracing algorithms. In An Introduction to Ray

Tracing, A. S. Glassner, Ed. Academic Press, San Diego, 1989, pp. 3377.

[109] HAINES, E. A. Beams o light: Confessions of a hacker. In SIGGRAPH

91 Frontiers in Rendering course notes. July 1991.

[110] HAINES, E. A. RONCHAMP: A case study for radiosity. In SIGGRAPH

91 Frontiers in Rendering course notes. July 1991.

[111] HAINES, E,. A., AND GREENBERG D. P. The light buffer: A shadow-
testing accelerator. IEEE Computer Graphics and Applications 6:9 (Sept.
1986), pp. 616.

[112] HAINES, E. A., AND WALLACE, J. R. Shaft culling for efficient

ray-traced radiosity. In Second Eurographics Workshop on Rendering
(Barcelona, Spain, May 1991).

[113] HALL, D. E., AND RUSHMEIER, H. Improved explicit radiosity method

for calculating non-lambertian reflections. The Visual Computer (to ap-
pear) 9:5 (1993), pp. 278288.

[114] HALL, R. Illumination and Color in Computer Generated Imagery.

Springer-Verlag, New York, 1989.

[115] HANRAHAN, P., AND KRUEGER, W. Reflection from layered sur-

faces due to subsurface scattering. Computer Graphics (SIGGRAPH 93
Proceedings) 27 (Aug. 1993).

[116] HANRAHAN, P., SALZMAN, D., AND AUPPERLE, L. A rapid hierarchi-

cal radiosity algorithm. Computer Graphics (SIGGRAPH 91 Proceedings)
25:4 (July 1991), pp. 197206.

Radiosity and Realistic Image Synthesis 358

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[117] HANRAHAN, P. M. Creating volume models from edge-vertex graphs. Com-

puter Graphics (SIGGRAPH 82 Proceedings) 16:3 (July 1982),
pp. 7784.

[118] HE, X. D., TORRANCE, K. E., SILLION, F. X., AND GREENBERG,

D. P. A comprehensive physical model for light reflection. Computer
Graphics (SIGGRAPH 91 Proceedings) 25:4 (July 1991), pp. 175186.

[119] HECKBERT, P. Adaptive radiosity textures for bidirectional ray trac-

ing. Computer Graphics (SIGGRAPH 90 Proceedings) 24:4 (Aug. 1 990),
pp 145154.

[120] HECKBERT, P. Simulating Global Illumination Using Adaptive Meshing.

PhD thesis, CS Division (EECS), University of California, Berkeley, June
1991.

[121] HECKBERT, P. Discontinuity meshing for radiosity. In Third Eurograph-

ics Workshop on Rendering (Bristol, UK, May 1992), pp. 203226.

[122] HECKBERT, P. Radiosity in flatland. Computer Graphics Forum (Euro-

graphics 92) 11:3 (Sept. 1992), pp. 181192.

[123] HECKBERT, P. S. Simulating Global Illumination Using Adaptive Mesh-

ing. PhD thesis, University of California, Berkeley, 1991.

[124] HECKBERT, P. S., AND HANRAHAN, P. Beam tracing polygonal ob-

jects. Computer Graphics (SIGGRAPH 84 Proceedings) 18:3 (July 1984),
pp. 119127.

[125] HECKBERT, P. S., AND WINGET, J. M. Finite element methods for

global illumination. Tech. Rep. UCP/CSD 91/643, Computer Science
Division (EECS), University of California, Berkeley, July 1991.

[126] HERMAN, R. A. A Treatise on Geometrical Optics. Cambridge Univer-

sity Press, 1900.

[127] HEWITT, H., AND VAUSE, A. S., Eds. Lamps and Lighting. American
Elsevier, New York, 1964.

[128] HORN, B. K. P. Robot Vision. MIT Press, 1986.

[129] HORN, B. K. P., Ed. Shape from Shading. MIT Press, 1989.

[130] HOTTEL, H. C., AND SAROfi M, A. F. Radiative Transfer. McGraw

Hill, New York, 1967.

Radiosity and Realistic Image Synthesis 359

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[131] HOWELL, J. R. A Catalog of Radiation Configuration Factors. McGraw

Hill, New York, 1982.

[132] I MMEL D. S., COHEN, M., AND GREENBERG, D. P. A radiosity

method for non-diffuse environments. Computer Graphics (SIGGRAPH
86 Proceedings) 20:4 (Aug. 1986), pp. 133142.

[133] JOHNSON, K. R., CURRAN, A. R., AND GONDA, T. G. Develop-

ment of a signature supercode. In Technical Proceedings 1938: Advances
in Senors, Radiometric Calibration, and Processing of Remotely Sensed
Data (SPIE International Symposium on Optical .Engineering and Photon-
ics in Aerospace and Remote Sensing) (Orlando, FL, Apr. 1993).

[134] KAJIYA, J. T. Anisotropic reflection models. In Computer Graphics

(SIGGRAPH 85 Proceedings) (July 1985), Vol. 19, pp. 1521.

[135] KAJIYA, J. T. The rendering equation. Computer Graphics (SIGGRAPH

86 Proceedings) 20:4 (Aug. 1986), pp. 143150.

[136] KAJIYA, J. T., AND HERZEN, B. P. V. Ray tracing volume densi-

ties. Computer Graphics (SIGGRAPH 84 Proceedings) 18:3 (July 1984),
pp. 165174.

[137] KAWAI , J., PAINTER, J., AND COHEN, M. Radioptimization: Goal-

based rendering. Computer Graphics (SIGGRAPH 93 Proceedings) 27
(Aug 1993).

[138] KINCAID, D., AND C HENEY, W. Numerical Analysis. Brooks/Cole,

1991.

[139] KIRK, D., AND VOORHIES , D. The rendering architecture of the

DN10000VS. Computer Graphics (SIGGRAPH 90 Proceedings) 24:4
(Aug 1990), pp. 299308.

[140] KIRK, D. B., AND ARVO, J. R. Unbiased sampling techniques for

image synthesis. Computer Graphics (SIGGRAPH 91 Proceedings) 25:4
(July 1991), pp. 153156.

[141] KOK, A. J. F., AND JANSEN, F. W. Adaptive sampling of area light

sources in ray tracing including diffuse interreflection. Computer Graphics
Forum (Eurographics 92) 11:3 (Sept. 1992), pp. 289298.

[142] KOK, A. J. F., YILMAZ, C., AND B IERENS, L. H. J. A two-pass ra-

diosity method for Bzier patches. In Photorealism in Computer Graphics

Radiosity and Realistic Image Synthesis 360

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

(Proceedings Eurographics Workshop on Photosimulation, Realism and

Physics in Computer Graphics, Rennes, France, Jun. 1990), Springer-
Verlag, pp. 115124.

[143] LAMBERT. Photometria sive de mensura et gradibus luminis, colorum et

umbrae. 1760.

[144] LANGUENOU, E., AND TELLIER, P. Including physical light sources

and daylight in global illumination. In Third Eurographics Workshop on
Rendering (Bristol, UK, May 1992), pp. 217225.

[145] LAUR, D., AND HANRAHAN, P. Hierarchical splatting: A progressive

refinement algorithm for volume rendering. Computer Graphics (SIG-
GRAPH 91 Proceedings) 25:4 (jul 1991), pp. 285288.

[146] LE, K. H. Finite element mesh generation methods: A review and

classification. Computer-Aided Design 20 (1988), pp. 2738.

[147] LEE, M., R EDNER, R., AND U SELTON, S. Statistically optimized

sampling for distributed ray tracing. Computer Graphics (SIGGRAPH 85
Proceedings) 19:3 (1985), pp. 6167.

[148] LESAEC, B., AND SCHLICK, C. A progressive ray-tracing-based ra-

diosity with general reflectance functions. In Photorealism in Computer
Graphics (Proceedings Eurographics Workshop on Photosimulation, Re-
alism and Physics in Computer Graphics, Rennes, France, Jun. 1990),
K. Bouatouch and C. Bouville, Eds., Springer-Verlag, pp. 101114.

[149] LEVOY, M. Display of surfaces from volume data. IEEE Computer

Graphics and Applications (May 1988), pp. 2937.

[150] LEVOY, M. Efficient ray tracing of volume data. ACM Transactions on

Graphics 9:3 (July 1990), pp. 245261.

[151] LEWIN, L. Dilogarithm and Associated Functions. Macdonald, London,

1958.

[152] LEWINS, J. Importance, the Adjoint Function: The Physical Basis of

Variational and Perturbation Theory in Transport and Diffusion Problems.
Pergamon Press, New York, 1965.

[153] LISCHINSKI, D., TAMPIERI, F., AND GREENBERG, D. P. Improving

sampling and reconstruction techniques for radiosity. Tech. Rep. TR 91
1202, Program of Computer Graphics, Cornell University, Aug. 1991.

Radiosity and Realistic Image Synthesis 361

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[154] LISCHINSKI, D., TAMPIERI, F., AND GREENBERG, D. P. Discon-

tinuity meshing for accurate radiosity. IEEE Computer Graphics and
Applications 12:6 (Nov. 1992), pp. 2539.
[155] LOVE, T. J. Radiative Heat Transfer. Merrill Publishing Company,
l968.
[156] MACADAM, D. L. Sources of Color Science. MIT Press, Cambridge,
MA, 1970.
[157] MALLEY, T. J. A shading method for computer generated images.
Masters thesis, Dept. of Computer Science, University of Utah, June
1988.
[158] MNTYL, M. An Introduction to Solid Modeling. Computer Science
Press, Rockville, md, 1988.
[159] MNTYL, M., AND SULONEN, R,. Gwb - a solid modeler with Euler
operators. IEEE Computer Graphics and Applications 2:7 (Sept. 1982),
pp. 1731.
[160] MARKS, J., WALSH, R., C HRISTENSEN, J., AND FRIEDELL , M. Im-
age and intervisibility coherence in rendering. In Proceedings of Graphics
Interface 90 (Toronto, Ontario, May 1990), Canadian Information Pro-
cessing Society, pp. 1730.
[161] MAX, N. Smooth appearance for polygonal surfaces. The Visual Com-
puter 5:3 (1989), pp. 160173.
[162] MAX, N. Optimal hemicube sampling. In 1993 Eurographics Rendering
Workshop (Paris, 1993).
[163] MAX, N. L., AND A LLISON, M. J. Linear radiosity approximation
using vertex-to-vertex form factors. In Graphics Gems III, D. Kirk, Ed.
AP Professional, Boston, 1992, pp. 318323.
[164] MAXWELL G. M., B AILEY, M. J., AND G OLDSCHMIDT, V. W. Cal-
culations of the radiation configuration factor using ray casting. Computer-
Aided Design 18:7 (Sept. 1986), pp. 371379.
[165] METAXAS, D., AND MILIOS, E. Color image reconstruction from
nonuniform sparse samples. In Eurographics 90 (1990), pp. 7586.
[166] MEYER, G. W. Color Calculations For and Perceptual Assessment of
Computer Graphic Images. PhD thesis, Program of Computer Graphics,
Cornell University, 1986.

Radiosity and Realistic Image Synthesis 362

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[167] MEYER, G. W., R USHMEIER, H. E., COHEN, M. F., GREENBERG,

D. P., AND TORRANCE, K. E. An experimental evaluation of computer
graphics imagery. ACM Transactions on Graphics 5:1 (Jan. 1986), pp. 30
50.

[168] MITCHELL, D., AND HANRAHAN, P. Illumination from curved reflec-

tors. Computer Graphics (SIGGRAPH 92 Proceedings) 26:4 (July 1992),
pp 283291.

[169] MOON, P. The Scientific Basis of Illumination Engineering. McGraw

Hill, New York, 1936.

[170] MORTENSON, M. E. Geometric Modeling. John Wiley & Sons, New

York, 1985.

[171] NAYLOR, B. F. Binary space partitioning trees: An alternative represen-

tation of polytopes. Computer-Aided Design 22:2 (Mar. 1990), pp. 250
253.

[172] NEUMANN, L., AND KELEMEN, C. Solution of interreflection problem

for very complex environments by transillumination method. In Second
Eurographics Workshop on Rendering (Barcelona, Spain, May 1991).

[173] NEWMAN , W., AND SPROULL, R. Principles of Interactive Computer

Graphics. McGraw-Hill, 1979.

[174] NICODEMUS, F. E., R ICHMOND, J. C., H SIA, J. J., GINSBERG ,

I. W., AND LIMPERIS, T. Geometric Considerations and Nomenclature
for Reflectance, NBS Monograph 160. National Bureau of Standards,
1977.

[175] NISHITA, T., AND NAKAMAE, E. Continuous tone representation

of three-dimensional objects taking account of shadows and interreflec-
tion. Computer Graphics (SIGGRAPH 85 Proceedings) 19:3 (July 1985),
pp. 2330.

[176] NISHITA, T., AND NAKAMAE, E. Continuous tone representation of

three-dimensional objects illuminated by sky light. Computer Graphics
(SIGGRAPH 86 Proceedings) 20:4 (Aug. 1986), pp. 125132.

[177] OPTICAL S OCIETY OF AMERICA COMMITTEE ON C OLORIMETRY

The Science of Color. Optical Society of America, Washington, DC,
1973.

Radiosity and Realistic Image Synthesis 363

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[178] PARKER, A., CHRISTOU, C., C UMMING, B., AND ZISSERMAN, A.

Evaluation of a radiosity-based method for generating images of 3-d
shapes. Perception 21:(Supplement 2)(1992), p. 18.

[179] PARKER, A. J., AND ZISSERMAN , A. The analysis of 3d

shape: Psychophysical principles and neural mechanisms. In Approaches
to Understanding Vision, B.W. Humphries, Ed. Blackwell, 1992, pp. 143
179.

[180] PEARSON, C. E., Ed. Handbook of Applied Mathematics. Van Nostrand

Reinhold, New York, 1990.

[181] PHONG, B. T. Illumination for computer-generated images. Tech. rep.,

Dept. of Computer Science, University of Utah, Salt Lake City, 1973

[182] PHONG, B. T. Illumination for computer generated pictures. Commu-

nications of the ACM 18:6 (1975), pp. 311317.

[183] POWELL, M., AND SABIN, M. Piecewise quadratic approximation

on triangles. ACM Transactions on Mathematical Software (Dec. 1977),
pp. 316325.

[184] PREPARATA, F. P., AND SHAMOS, M. I. Computational Geometry:

An Introduction. Springer-Verlag, New York, 1985.

[185] PRESS , W., TEUKOLSKI, S., V ETTERLING, W., AND FLANNERY,

B. Numerical Recipes in C, The Art of Scientific Computing, 2nd Edition.
Cambridge University Press, Cambridge, 1992.

[186] PUECH, C., SILLION, F., AND VEDEL., C. Improving interaction with
radiosity-based lighting simulation programs. Computer Graphics (Pro-
ceedings of the 1990 Symposium on Interactive 3D Graphics) 24:2 (Mar.
1990), pp. 5157.

[187] PUEYO, X. Diffuse interreflection. techniques for form-factor compu-

tation: A survey. The Visual Computer 7:4 (July 1991), pp. 200209.

[188] PURGATHOFER, W., AND ZEILLER, M. Fast radiosity by paralleliza-

tion. In Photorealism in Computer Graphics (Proceedings Eurographics
Workshop on Photosimulation, Realism and Physics in Computer Graph-
ics, Rennes, France, Jun. 1990), Springer-Verlag, pp. 171181.

[189] RATLIFF, F. Mach Bands: Quantitative Studies on Neural Networks in

the Retina. Holden-Day, Inc., San Francisco, 1965.

Radiosity and Realistic Image Synthesis 364

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[190] RECKER, R. J. Improved techniques for progressive refinement radios-

ity. Masters thesis, Program of Computer Graphics, Cornell University,
Jan. 1990.

[191] RECKER, R. J., GEORGE, D. W., AND GREENBERG, D. P. Acceler-

ation techniques for progressive refinement radiosity. Computer Graphics
(1990 Symposium on Interactive 3D Graphics) 24:2 (Mar. 1990), pp. 59 66.

[192] REICHERT, M. C. A two-pass radiosity method driven by lights and

viewer position. Masters thesis, Program of Computer Graphics, Cornell
University, Jan. 1992.

[193] REWALD, J. The History of Impressionism. Museum of Modern Art,

New York, 1973.

[194] ROCKWOOD , A., HEATON, K., AND D AVIS, T. Real-time rendering

of trimmed surfaces. Computer Graphics (SIGGRAPH 89 Proceedings)
23:3 (July 1989), pp. 107116.

[195] ROGERS, D. Procedural Elements for Computer Graphics. McGraw-Hill,

1985.

[196] RUSHMEIER , H., PATTERSON, C., AND VEERASAMY, A. Geometric

simplification for indirect illumination calculations. In Graphics Interface
93 Proceedings (Toronto, May 1993).

[197] RUSHMEIER , H. E. Extending the radiosity method to transmitting

and specularly reflecting surfaces. Masters thesis, Program of Computer
Graphics, Cornell University, 1986.

[198] RUSHMEIER, H. E. Realistic Image Synthesis for Scenes with Radiatively

Participating Media. PhD thesis, Program of Computer Graphics, Cornell
University, 1988.

[199] RUSHMEIER, H. E. Radiosity input/output. In Radiosity, SIGGRAPH

92 course notes, Vol. 11. ACM Press, July 1992, pp. 152168.

[200] RUSHMEIER , H. E., and torrance, K. E. The zonal method for cal-
culating light intensities in the presence of a participating medium. Com-
puter Graphics (SIGGRAPH 87 Proceedings) 21:4 (July 1987), pp. 293
302.

Radiosity and Realistic Image Synthesis 365

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[201] RUSHMEIER , H. E., AND TORRANCE, K. E. Extending the radiosity

method to include specularly reflecting and translucent materials. ACM
Transactions on Graphics 9:1 (Jan. 1990), pp. 127.

[202] SALA, A. Radiant Properties of Materials. Elsevier, Amsterdam, 1986.

[203] SALESIN , D., LISCHINSKI, D., AND DEROSE, T. Reconstructing il-

lumination functions with selected discontinuities. In Third Eurographics
Workshop in Rendering (Bristol, UK, May 1992), pp. 99112.

[204] SAMET, H. Design and Analysis of Spatial Data Structures. Addison-

Wesley, Reading, MA, 1990.

[205] SCHOENBERG, I. J. Contributions to the problem of the approxima-

tion of equidistant data by analytic functions. Quarterly Applications of
Mathematics 4 (1946), pp. 4599.

[206] SCHRDER, P., AND HANRAHAN, P. A closed form expression for the
form factor between two polygons. Tech. Rep. Cs-40493, Department
of Computer Science, Princeton University, Jan. 1993.

[207] SCHUMAKER, L. L. Triangulations in CAGD. IEEE Computer Graphic.

and Applications 13:1 (Jan. 1993), pp. 4752.

[208] SCHUMAKER, R. B., BRAND, M. G., AND SHARP, W. Study for

applying computer-generated images to visual simulation, AFHRL-TR-
6914. Tech. rep., U.S. Air Force Human Resources Lab, 1969.

[209] SEGAL, M. Using tolerances to guarantee valid polyhedral modeling

results. Computer Graphics (SIGGRAPH 90 Proceedings) 24:4 (Aug.
1990), pp. 105114.

[210] SHAO, M.-Z., PENG, Q.-S., AND LIANG, Y.-D. A new radiosity ap-
proach by procedural refinements for realistic image synthesis. Computer
Graphics (SIGGRAPH 88 Proceedings) 22:4 (Aug. 1988), pp. 93101.

[211] SHEPHARD, M. S. Approaches to the automatic generation and control

of finite element meshes. Applied Mechanics Review 41:4 (Apr. 1988),
pp. 169185.

[212] SHIRLEY, P. Physically Based Lighting Calculations for Computer

Graphics. PhD thesis, Dept. of Computer Science, U. of Illinois, Urbana-
Champaign, Nov. 1990.

Radiosity and Realistic Image Synthesis 366

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[213] SHIRLEY, P. A ray tracing method for illumination calculation in diffuse-

specular scenes. In Proceedings of Graphics Interface 90 (Toronto, On-
tario, May 1990), Canadian Information Processing Society, pp. 205212.

[214] SHIRLEY, P. Radiosity via ray tracing. In Graphics Gems II, J. Arvo,
Ed. AP Professional, Boston, 1991, pp. 306310.

[215] SHIRLEY, P., AND WANG, C. Direct lighting calculation by Monte

Carlo integration. In Second Eurographics Workshop on Rendering
(Barcelona, Spain, May 1991).

[216] SIEGEL, R., AND HOWELL, J. R. Thermal Radiation Heat Transfer,

3rd Edition. Hemisphere Publishing Corporation, New York, 1992.

[217] SILLION , F., ARVO, J. R., WESTIN , S. H., AND GREENBERG, D. P.

A global illumination solution for general reflectance distributions. Com-
puter Graphics (SIGGRAPH 91 Proceedings) 25:4 (July 1991), pp. 187
196.

[218] SILLION , F., AND P UECH, C. A general two-pass method integrat-

ing specular and diffuse reflection. Computer Graphics (SIGGRAPH 89
Proceedings) 23:3 (July 1989), pp. 335344.

[219] SMITH, F. W., Ed. Crc Handbook of Marine Science. Crc Press,
Cleveland, Ohio, 1974.

[220] SMITS , B. E., A RVO, J. R., AND SALESIN, D. H. An importance-

driven radiosity algorithm. Computer Graphics (SIGGRAPH 92 Proceed-
ings) 26:4 (July 1992), pp. 273282.

[221] SPARROW, E. A new and simpler formulation for radiative angle factors.
Transactions of the ASME, The Journal of Heat Transfer 85:2 (1963),
pp. 8188.

[222] SPARROW, E., AND C ESS, R. Radiation Heat Transfer. Hemisphere

Publishing Corporation, Washington, 1978.

[223] SPENCER, S. N. The hemisphere radiosity method: A tale of two algo-

rithms. In Photorealism in Computer Graphics (Proceedings Eurographics
Workshop on Photosimulation, Realism and Physics in Computer Graph-
ics, 1990) (1991), K. Bouatouch and C. Bouville, Eds., pp. 12735.

[224] SPIVAK, M. Calculus on Manifolds. Benjamin/Cummings, 1965.

Radiosity and Realistic Image Synthesis 367

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[225] STEVENS, S. S., AND STEVENS, J. C. Brightness function: Paramet-

ric effects of adaptation and contrast. Journal of the Optical Society of
America 53:11 (Nov. 1960), pp. 1139.
[226] STOER, J., AND BULIRSCH, R. Introduction to Numerical Analysis.
Springer-Verlag, New York, 1980.
[227] STURZLINGER, W. Radiosity with voronoi-diagrams. In Third Euro-
graphics Workshop on Rendering (Bristol, UK, May 1992), pp. 169177.
[228] SUTHERLAND, I. E. Sketchpad: A man-machine graphical communi-
cation system. In Proceedings of the Spring Joint Computer Conference
(1963).
[229] TAKAGI, A., TAKAOKA, H. OSHIMA, T., AND OGATA, Y. Accurate
rendering technique based on colorimetric conception. Computer Graphics
(SIGGRAPH 90 Proceedings) 24:4 (Aug. 1990), pp.263272.
[230] TAMPIERI, F. Accurate form-factor computation. In Graphics Gems III,
D. Kirk, Ed. AP Professional, Boston, 1992, pp. 329333.
[231] TAMPIERI, F., AND LISCHINSKI, D. The constant radiosity assumption
syndrome. In Second Eurographics Workshop on Rendering (Barcelona,
Spain, May 1991).
[232] TELLER, S., AND HANRAHAN, P. Global visibility algorithms for il-
lumination computations. Computer Graphics (SIGGRAPH 93 Proceed-
ings) 27 (Aug. 1993)
[233] TELLER, S. J. Computing the antipenumbra of an area light. Computer
Graphics (SIGGRAPH 92 Proceedings) 26:4 (July 1992), pp. 139148.
[234] TELLER, S. J., AND SEQUIN, C. H. Visibility preprocessing for inter-
active walkthroughs. Computer Graphics (SIGGRAPH 91 Proceedings)
25:4 (July 1991), pp. 6169.
[235] THOMPSON, J. F., W ARSI, Z., AND MASTIN, C. W. Numerical Grid
Generation. North-Holland, New York, 1985.
[236] TORRANCE, K. E., AND SPARROW, E. M. Theory for off-specular
reflection from roughened surfaces. Journal of the Optical Society of
America 57:9 (Sept. 1967), pp. 11051114.
[237] TOULOUKIAN, Y. S. AND D EWITT, D. P. Thermophysical Properties
of Matter, Vols. 7, 8: Thermal Radiative Properties. IFI/Plenum, New
York, 1972.

Radiosity and Realistic Image Synthesis 368

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[238] TUMBLIN, J., AND R USHMEIER, H. E. Tone reproduction for realistic

computer generated images. Tech. Report GI GVU-9113, Graphics, Vi-
sualization & Usability Center, College of Computing, Georgia Institute
of Technology, 1991.

[239] VAN LIERE, R. Divide and conquer radiosity. In Second Eurographics

Workshop on Rendering (Barcelona, Spain, May 1991).

[240] VARSHNEY, A., AND PRINS , J. F. An environment-projection ap-

proach to radiosity for mesh-connected computers. In Third Eurographics
Workshop on Rendering (Bristol, UK, May 1992), pp. 271281.

[241] VEDEL, C. Improved storage and reconstruction of light intensities on

surfaces. In Third Eurographics Workshop on Rendering (Bristol, UK,
May 1992), pp. 113121.

[242] VEDEL, C., AND P UECH, C. A testbed for adaptive subdivision in

progressive radiosity. In Second Eurographics Workshop on Rendering
(Barcelona, Spain, May 1991).

[243] VERBECK, C. P., AND GREENBERG, D. P. A comprehensive light-

source description for computer graphics. IEEE Computer Graphics and
Applications 4:7 (July 1984), pp. 6675.

[244] VILAPLANA, J. Parallel radiosity solutions based on partial result mes-

sages. In Third Eurographics Workshop on Rendering (Bristol, UK, May
1992), pp. 259270.

[245] VILAPLANA, J., AND PUEYO, X. Exploiting coherence for clipping and
view transformations in radiosity algorithms. In Photorealism in Computer
Graphics (Proceedings Eurographics Workshop on Photosimulation, Re-
alism and Physics in Computer Graphics, Rennes, France, Jun. 1990),
K. Bouatouch and C. Bouville, Eds., Springer-Verlag, pp. 137150.

[246] W ALLACE, J. R., COHEN, M. F., AND GREENBERG, D. P. A

two-pass solution to the rendering equation: A synthesis of ray tracing
and radiosity methods. Computer Graphics (SIGGRAPH 87 Proceedings)
21:4 (July 1987), pp. 311320.

[247] W ALLACE, J. R., ELMQUIST, K. A., AND HAINES , E. A. A ray trac-

ing algorithm for progressive radiosity. Computer Graphics (SIGGRAPH
89 Proceedings) 23:3 (July 1989), pp. 315324.

Radiosity and Realistic Image Synthesis 369

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[248] WANG, C. Physically correct direct lighting for distribution ray tracing.
In Graphics Gems III, D. Kirk, Ed. AP Professional, Boston,, 1992,
pp. 307313.

[249] W ANG, Y., AND DAVIS, W. A. Octant priority for radiosity image
rendering. In Proceedings of Graphics Interface 90 (Toronto, Ontario,
May 1990), Canadian Information Processing Society, pp. 8391.

[250] WANGER., L. The effect of shadow quality on the perception of spa-

tial relationships in computer generated imagery. In Computer Graphics,
Special Issue, (Proceedings 1992 Symposium on Interactive 3D Graphics)
(Cambridge, Mass., Mar. 1992), ACM Press, pp. 3942.

[251] WARD, G. Evaluating a real lighting simulation. In Radiosity, SIG-

GRAPH 90 course notes, Vol. 21. ACM Press, Aug. 1990.

[252] WARD, G. The radiance lighting simulation system. In Global Illumina-

tion, SIGGRAPH 92 course notes, Vol. 18. ACM, July 1992.

[253] WARD, G. J. Measuring and modeling anisotropic reflection. Computer

Graphics (SIGGRAPH 92 Proceedings) 26:2 (July 1992), pp.265272.

[254] W ARD, G. J., RUBINSTEIN, F. M., AND CLEAR, R. D. A ray tracing

solution for diffuse interreflection. Computer Graphics (SIGGRAPH 88
Proceedings) 22:4 (Aug. 1988), pp. 8592.

[255] WARNOCK, J. A hidden-surface algorithm for computer generated half-

tone pictures. Technical Report TR 4-15, Dept. of Computer Science,
University of Utah, June 1969.

[256] WATSON, D. Computing the n-dimensional Delaunay tesselation with

application to Voronoi polytopes. The Computer Journal 24:2 (1981),
pp. 167172.

[257] WATSON, D. F. Contouring: A Guide to the Analysis and Display of

Spatial Data. Pergamon Press, New York, 1982.

[258] WATT, A. Fundamentals of Three-Dimensional Computer Graphics.

Addison-Wesley, 1989.

[259] WATT, M. Light-water interaction using backward beam tracing. Com-

puter Graphics (SIGGRAPH 90 Proceedings) 24:4 (Aug. 1990), pp. 377
385.

Radiosity and Realistic Image Synthesis 370

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[260] W EGHORST, H., HOOPER, G. J., AND GREENBERG, D. P. Improved

computational methods for ray tracing. ACM Transactions on Graphics
3:1 (Jan. 1984), pp. 5269.

[261] WEILER, K. Edge-based data structures for solid modeling in curved-

surface environments. IEEE Computer Graphics and Applications 3:1
(Jan. 1985), pp. 2140.

[262] WEILER, K. Topological Structures for Geometric Modeling. PhD the-

sis, Computer and Systems Engineering, Rensselaer Polytechnic Institute,
Troy, New York, Aug. 1986.

[263] W ESTIN, S. H., ARVO, J. R., AND TORRANCE, K. E. Predicting

reflectance functions from complex surfaces. Computer Graphics (SIG-
GRAPH 92 Proceedings) 26:2 (July 1992), pp. 255264.

[264] WESTOVER, L. Footprint evaluation for volume rendering. Computer

Graphics (SIGGRAPH 90 Proceedings) 24:4 (July 1990), pp. 367376.

[265] W HITTED, T. An improved illumination model for shaded display.

Communications of the ACM 23:6 (1980), pp. 343349.

[266] W ILLIAMS, L. Pyramidal parametrics. In Computer Graphics (SIG-

GRAPH 83 Proceedings) (July 1983), Vol. 17, pp. 111.

[267] WILSON, P. R. Euler formulas and geometric modeling. IEEE Computer

Graphics and Applications 5:8 (Aug. 1 985), pp. 2436.

[268] WOO, T. A combinatorial analysis of boundary data structure schemata.

IEEE Computer Graphics and Applications 5:3 (Mar. 1985), pp. 1927.

[269] XU, H., P ENG, Q.-S., AND LIANG, Y.-D. Accelerated radiosity
method for complex environments. In Visual Computing: Integrating
Computer Graphics with Computer vision (Proceedings of CG Interna-
tional 92) (Tokyo, Sept. 1989), Springer-Verlag, pp. 895905.

[270] ZATZ, H. R. Galerkin radiosity: A higher order solution method for

global illumination. Masters thesis, Program of Computer Graphics, Cor-
nell University, Aug. 1992.

[271] ZHANG, N. Two methods for speeding up form-factor calculation. In

Second Eurographics Workshop on Rendering (Barcelona, Spain, May
1991).

Radiosity and Realistic Image Synthesis 371

Edited by Michael F. Cohen and John R. Wallace
 BIBLIOGRAPHY

[272] ZHU, Y., PENG, Q. AND LIANG, Y. Peris: a programming envi-

ronment for realistic image synthesis. Computers and Graphics 12:3/4
(1988), pp. 299307.

[273] ZIENKIEWICZ, O. C. The Finite Element Method, 4th Edition. McGraw-

Hill, London, 1989.

[274] ZISSERMAN, A., GIBLIN, P., AND B LAKE, A. The information avail-
able to a moving observer from specularities. Image and Vision Computing
7:1 (1989), pp. 3842.

Radiosity and Realistic Image Synthesis 372

Edited by Michael F. Cohen and John R. Wallace
 INDEX

Index

A posteriori mesh 154 Bidirectional ray tracing 304, 322

limitations 222 Bidirectional reflection distribution
A priori mesh 154 function 28
Adaptive subdivision 157, 169, 217, Boundary element method 46
224 BRDF 28
BSP-tree 166 components 299
quadtree 166 data 324, 344
triquadtree 214 remote sensing 338
using templates 214 spherical harmonics 313
Advancing front 218 Brightness 269
Ambient 122 BSP-tree 217, 238
Anisotropy 29 2D 166, 231
Antialiasing 244 3D 231
Architectural design 332 balanced 218
Aspect ratio 145, 218 shadow volumes 230
B 25 B-spline 257
Barycentric coordinates 52, 247 BTDF 317
Basis function 10, 46 Bump mapping 266, 267
bilinear 52, 248 C 143
box 49 Candela 27
constant 49, 57, 244 Cathode ray tube (CRT) 3
Haar basis 190 Caustic 304
hat 49 Clough-Tocher element 144
hierarchical 167, 187 Color bleeding 341
higher order 52, 60 Color 41, 109, 267, 273
linear 49, 144 CIE XYZ space 278, 282
mesh 131 emission spectra 298
order 142 gamut 3
orthonormal 60 luminous efficiency function 275
rendering 244, 245 matching functions 276
spherical harmonics 313 metamers 274
support 4748 monitor 267
wavelet 190 perception 274-275
Bzier patch 252253 RGB space 282
quadratic triangle 253 sampling 280

Radiosity and Realistic Image Synthesis 373

Edited by Michael F. Cohen and John R. Wallace
 INDEX

spectral sampling 283 concave 145, 251

transformations 279 conformance 147, 152, 214, 218
Conformance 147, 152, 214, 218 constant 49, 244246, 262
Constant elements 262 continiuty 244
Continuity 143 hermite 144
Convexity 63 isoparametric 144
Coordinate Lagrange 143
barycentric 52 linear 49, 143, 245246, 256
Critical surface 226 master 61
CRT 268, 279 order 142
Delaunay triangulation 217221,246 orientation 146
constrained 220 parametric coordinates 247
Delta form factor 83 quadratic 253
nonconstant bases 99 shape 144
Delta function 31 size 139
Diagonally dominant 110 standard 52
Differential form factor 67 triangular 247
Diffuse reflection 32 Emitted energy 26, 40
Directional diffuse reflection 34 Error analysis 343
Discontinuity 139, 149, 164 Error estimate
derivative 150, 224228 gradient-based 160
first derivative 224228, 244 heuristic 160
shape perception 339 higher order 161
value 147, 150152, 222-224 min-max search 164
Discontinuity meshing 154, 164, 222- residual 162
233, 253, 259, 345 Error metric 48, 53
critical surface 226 finite 54
reconstruction 256 function norm 133
shadow volumes 229 image importance 201
value discontinuities 222 image-based 135, 258, 344
value 139 kernel-based 135
Dynamic environments 126 local estimate 132
geometry 127 oracle function 178186, 201
lighting 126 perceptually-based 136, 343
reflectivity 127 residual 134
E 40 true error l32, l34
Element 8, 46, 48 view-dependent 201
aspect ratio 145, 218 Experimental validation 340
bilinear 52 Exposure 17, 23
C1 252 Extended form factor 307, 323
Clough-Tocher 1 44, 25 3 transmission 3l8

Radiosity and Realistic Image Synthesis 374

Edited by Michael F. Cohen and John R. Wallace
 INDEX

Fine arts 340 visibility 68

Finite element method 8, 45 volume-to-surface 328
Galerkin 56 volume-to-volume 328
history 8 Form factor algorithms
point collocation 55 acceleration 103
steps 46 area sampling 90
weighted residuals 56 contour integral 95
Flatland radiosity 172 disk approximation 92
Flux Phi 19 hemicube 80
Fog 325 Malleys method 90
Form factor 47, 58, 6566, 167, 172 Monte Carlo 8990
algebra 72 numerical solutions 75
aliasing 84, 89 simple shapes 72
area-to-area 69 single plane 88
area-to-hemisphere 69 Fresnel formula 35
closed form 71 Function norm 133
contour integral 70 Function space 42
differential 39, 67 dimension 42
disk approximation 146 finite 42, 52
element shape 146 Function subspace 42
error metric 135, 183 Function
extended 307, 318, 323 continuity 143
general BRDF 312, 315 projection 42, 47, 196
geometry 68 G(x, x) 3941
history 70 Galerkin 56
matrix qualities 110 constant element 57
Monte Carlo 94, 262 Gamma correction 268
non-area light sources 291 Gathering 115
nonconstant bases 98 super-shoot-gather 125
Nusselt analog 80 Gauss-Seidel 114
occlusion testing 223 algorithm 115
per pixel 260, 262 Geometric decomposition 216
point-to-polygon 72 Global cube 309
polygon-to-polygon 74 Global illumination 6, 38
quadrature 77, 94 Glossy reflection 33, 299, 313
ray traced 318 Goniometric diagram 292293
reciprocity 68, 92 Goniometric 26
sampling artifacts 262 Gouraud shading l44, 249251
singularities 68, 100 Gradient 241
test environment 96 analytic 164
translucency 318 numerical differencing 164

Radiosity and Realistic Image Synthesis 375

Edited by Michael F. Cohen and John R. Wallace
 INDEX

tangent plane 164 Integral equation 40

Grid superposition 210 Intensity 25
H-refinement 155 Interpolation
Hardware rendering 249, 284 B-spline 253
specular highlights 284, 324 barycentric coordinates 247
texture mapping 285 Bzier 252
visibility preprocessing 286 bilinear 247248
Helmholtz reciprocity principle 29 C0 245
Hemicube 80 C1 252, 256
acceleration 103 Clough-Tocher 253
delta form factor 83 quadratic 252
heat transfer application 339 Irradiance 24
nonconstant bases 99 Isoparametric 62
resolution 85 Item buffer 82, 105
Shaos method 312 Iterative refinement 312
volume form factor 328 Jacobi iteration 182
Hierarchy 8, 167 Jacobi method 113
clustering 345 Joule 15, 27
geometric simplification 346 K (see also Matrix) 56
glossy reflection 316 adjoint operator 202
importance-based 205 coefficients 66
multilevel 176 components 65
patch subdivision 172 flatland radiosity 172
quadtree 176 Galelkin 57
spatial subdivision 346 matrix qualities 110
two-level 169, 175 matrix 171
Human perception point collocation 56
error metric 136 visualization 172
Illuminance 24 L 19
Image synthesis 2 Lagrange basis 143
goals 2 Lambertian diffuse 32
history 4 Leaf canopy simulation 338
limitations 2 Light leak 150
tractability 42 Light source
View-dependent 43 emission spectra 298
View-independent 44 goniometric diagram 293, 29
Importance meshing 201, 265 ideal diffuse assumption 341
Importance sampling 78 nondiffuse luminaire 293
Indirect illumination 38 nondiffuse 289, 336
Infrared signature analysis 339 normalization 297
Inner product 54 parallel 293

Radiosity and Realistic Image Synthesis 376

Edited by Michael F. Cohen and John R. Wallace
 INDEX

point 293 symmetry 110

sky light 295, 298 Matrix solution 109
spot light 295 gathering 181
time varying 336 Gauss-Seidel 114
Lighting design 334, 337 Jacobi iteration 182
Lighting optimization 337 shooting 181
Light 14 Southwell 116
coherent 14 Mesh 48
electromagnetic spectrum 14 a posteriori 154
flux 19 a priori 154
incoherent 14 artifacts 137
particles 18 aspect ratio 145
polarized 14 BSP-tree 217
power 17 boundary 211212
spectral sampling 284 conformance 147, 152, 214, 218
Linear equation solver continuity 142
direct 112 density 139
Gauss-Seidel 114 grading 147, 214
initial guess 1 13 hierarchy 238
iterative 112 nonuniform 141, 147
Jacobi 113 optimal 131
relaxation 113 quadtree 212, 214
Southwell 116 relaxation 155, 221
Local illumination 5, 37 shadow boundaries 163
Lumen 16, 27 smoothing 210, 221222
Luminance 19, 269 template 210
pixel 268 topology 218, 223, 231, 235
Luminosity 25 238
Luminous efficiency function 275 transition 145
Luminous intensity 25 uniformity 145
Lux 27 uniform 137139, 154, 166
M 60 user parameters 331, 345
Mach band 139, 143, 252253 Mesh topology 234238
Machine vision 339 adjacency graph 234
Master element 61 data structures 235
Matching functions 276 Euler operators 236
Matrix T-vertices 214218
condition 111 traversal queries 238
diagonal dominance 110 Meshing algorithms
sparsity 110 adaptive subdivision 217
spectral radius 111 advancing front 218

Radiosity and Realistic Image Synthesis 377

Edited by Michael F. Cohen and John R. Wallace
 INDEX

automatic 152 Participating media 318, 325

decomposition 216 anisotropic 330
Grid superposition 210 isotropic 326
multiblocking 212 phase function 326
nodes-elements-together 217 zonal method 327
nodes-first triangulation 219 Patch subdivision 172, 175
quadtree 211 Path tracing 319
recursive subdivision 212 Penumbra 149150, 226, 229
template mapping 211 Perception 267
Metamers 274 evaluation of images 341
Monte Carlo 42, 77, 89, 99 shape understanding 338
importance sampling 78 Phase function 326
quasi 42 Photometry 15
ray tracing 309, 346 history 15
Multi-pass method 265, 323 units 27
Multiblocking 212 Photorealism 2
Ni 42, 48 Pixel 5
N-body problem 177 luminance 268
Neumann series 111, 299 Pi 40
Nit 27 Point collocation 55
Nodal averaging 246 occlusion testing for 223
Node 46, 48 Progressive refinement 8, 119
Nonuniform mesh 141 ambient 122
Norm 54 general brdf 311, 314
function 133 overrelaxation 125
Numerical differencing 164 southwell iteration 120
Numerical integration (see also Quadra- Projected solid angle 24
ture) 76 Projection 42
Numerical integration 76 Pseudocode
adaptive ray shooting 314 directional radiance 315
Nusselt analog 80 GatherRadShootImp 208
Octree 241 GatherRad 181
Opera lighting 335 Gauss-Seidel 115
Optics 14 Hemicube 85, 86
Oracle function 178186, 201 HierarchicalRad 185
Overrelaxation 124 ImportanceDrivenRad 205
P-refinement 155, 157 Monte Carlo 78, 95
Parallel 129 Oracle1 183
fine grained 130 Oracle2 186
workstations 129 progressive refinement 120
Parametric mapping 61 PushPullRad 182

Radiosity and Realistic Image Synthesis 378

Edited by Michael F. Cohen and John R. Wallace
 INDEX

Refinelink 186 Raster graphics 5

Refine 179 Ray tracing 6, 43
SolveDualSystem 206 acceleration 106
SolveSystem 180 bidirectional 304, 322
Southwell iteration 118 for vision research 338
Quadratic elements 253 from the eye 302, 319
Quadrature 76 from the light 303
form factor 77 shaft-culling 106
gaussian 77 Reciprocity 29, 59
importance sampling 78 Recursive subdivision 212
Monte Carlo 77 Reflectance 31
Quadtree 171, 176, 211214, 233, biconical 31
238239 hemispherical 32
balanced or restricted 214, 217 Reflectance equation 30
tri-quadtree 214 Reflection frustum 320
33 Reflection 28
R-refinement 155157, 221 diffuse 32
RGB to XYZ 279 directional diffuse 34
RMS error 133 Fresnel 35
Radiance program 241, 335 glossy 33
Radiance 19 Lambertian 32
Radiant exitance 25 microfacet 35
Radiant intensity 25 mirror 30
Radiometry 15 rough surface 34
units 27 specular 33
Radiosity Regular expression 300
definition 25 Relaxation 113
diffuse assumption 40 Remeshing 1551 57
history 7 Remote sensing 338
Radiosity equation 40, 41 Rendering
assumptions 289 hardware 249
classical 59 Rendering equation 8, 36
matrix form 56, 59 definition 39
singularities 247 Residual function 47
translucency 318 Residual 54, 113
Radiosity function Shadow 149
discontinuity 164, 222 Shadow boundaries 163, 226, 229
gradient 163, 164 Shadow leak 150, 224
Radiosity gradient Shaft-culling 106
analytic 163 Shape function (see also Basis func-
Radiosity texture (Rex) 239 tion) 46

Radiosity and Realistic Image Synthesis 379

Edited by Michael F. Cohen and John R. Wallace
 INDEX

Shooting 117 Transport path 300

super-shoot-gather 125 Transport Theory 17
Singularities 100 Triangulation 166, 210, 214
Sky light 295 constrained Delaunay 220
spectral distribution 298 Delaunay 217221
Slave nodes 214215, 218 Gouraud shading 251
Smoke 325 nodes-first 219
Solid angle 20 Two point transport 38
projected 24 Two-pass method 259, 319, 322, 333
Southwell iteration 116 direct illumination 264
algorithm 117 Monte Carlo 260
residual update 117 Umbra 149150, 226, 229
Spectral luminous relative efficiency Uniform mesh 137139, 166
16 V(x, x ) 38
Spectral reflectance data 324 Vanishing moment 195
Specular reflection 33 View coherence 104
Spherical harmonics 313 View-dependent 43
volumes 330 View-dependent solution 201
Spot light 295 View-independent 7, 44
Steradian 20 Visibility
Stokes theorem 70 acceleration 103
Subdivision preprocessing 286
adaptive 157, 169 Z-buffer 82
BSP-tree 166 Visual events 226
image-based 240 critical surface 226
patch 172, 175 EEE 226, 233
quadtree 171 VE and EV 231
Super-shoot-gather 125 VE or EV 226
T-vertex 214218, 251 Visual shape understanding 338
Table of Terms 41 Volume rendering 329
Talbot 16, 27 Voronoi diagram 221, 246
Template mapping 211 Walkthrough 284, 332
Texture mapping 239, 266, 337338 Watt 15, 27
Theatrical lighting 335 Wavelet 190
Three-point transport 39, 316 Haar basis 190
Throughput 23 detail function 190
Tone reproduction 269 smooth function 190
Translucency 318 vanishing moment 195
Transmission 317 Weighted residual method 54
Transport paths 305 Winged-edge data structure 218, 223,
double counting 265, 324 231, 235238

Radiosity and Realistic Image Synthesis 380

Edited by Michael F. Cohen and John R. Wallace
 INDEX

Euler operators 236

performance 235
XYZ to RGB 279
Z-buffer 82, 105
Zonal method 327

Radiosity and Realistic Image Synthesis 381

Edited by Michael F. Cohen and John R. Wallace