See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/5507507

Image-Based Bidirectional Reflectance Distribution Function Measurement

Article  in  Applied Optics · July 2000

DOI: 10.1364/AO.39.002592 · Source: PubMed

CITATIONS READS

216 1,156

4 authors, including:

Stephen Westin
Grammatech Inc.
18 PUBLICATIONS   1,539 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Stephen Westin on 18 January 2017.

The user has requested enhancement of the downloaded file.

Image-based bidirectional reflectance distribution
function measurement

Stephen R. Marschner, Stephen H. Westin, Eric P. F. Lafortune, and

Kenneth E. Torrance

We present a new image-based process for measuring a surface’s bidirectional reflectance rapidly,
completely, and accurately. Requiring only two cameras, a light source, and a test sample of known
shape, our method generates densely spaced samples covering a large domain of illumination and
reflection directions. We verified our measurements both by tests of internal consistency and by com-
parison against measurements made with a gonioreflectometer. The resulting data show accuracy
rivaling that of custom-built dedicated instruments. © 2000 Optical Society of America
OCIS codes: 290.5820, 120.5820, 160.4760, 290.5880, 110.2960.

1. Introduction Several researchers have built BRDF measure-

One of the major goals of computer graphics is to ment instruments specifically for computer graphics.
synthesize realistic images of scenes modeled in a Some have used custom-made equipment; others
computer. To create such a realistic image, a com- have measured only a small part of the hemisphere to
plete model or description of reflectance is needed for supply data for fitting the parameters of a simple
each reflective surface in the scene. In optics termi- BRDF model. None has stressed absolute accuracy,
nology, the complete bidirectional reflectance distri- instead aiming for visually plausible reflectances.
bution function 共BRDF兲 described by Nicodemus et In this paper we present a system that measures
al.1 is needed. In contrast to many optical inspec- reflectance quickly, completely, and without special
tion and surface metrology applications, computer equipment. The method works by our taking a se-
rendering requires BRDF information over the com- ries of photographs of a curved object. Each image
plete scattering hemisphere for any incident direc- captures light reflected from many differently ori-
tion, including wavelength-dependent information ented parts of the surface, and with the knowledge of
for color. Computer graphics has long been aware of the sample shape and the light source position the
the BRDF literature, beginning with Blinn’s2 adap- photographs can be analyzed to determine the sam-
tation of the work of Torrance and Sparrow,3 but ple’s BRDF. By using a curved test sample and an
physically based models are still used only occasion- imaging detector, we eliminate the mechanism
ally, both because of their complexity and because needed to position the detector in a traditional go-
parameters are not readily available. Rendering nioreflectometer; and by using automated photo-
from measured BRDF data is also possible,4 but even grammetry to measure the source position, we
rarer because the necessary BRDF data are not eliminate the precise source positioning mechanism.
widely available. Although the apparatus is simple and the measure-
ment is rapid, the technique provides millions of ac-
curate samples that cover the full hemisphere to
When this research was undertaken, S. R. Marschner, S. H. angles near grazing.
Westin 共[email protected]兲, E. P. F. Lafortune, and K. E. Torrance
共[email protected]兲 were with the Program of Computer Graphics, 2. Prior Research
Cornell University, 580 Rhodes Hall, Ithaca, New York 14853- The BRDF is measured traditionally with a goniore-
3801. S. R. Marschner 共[email protected]兲 is now with Mi-
flectometer, positioning a source and detector with
crosoft Research, One Microsoft Way, Redmond, Washington
98052-6399. E. P. F. Lafortune 共[email protected]兲 is respect to a flat sample. Four degrees of rotational
now with Luciad, Parijsstraat 74, Leuven B-3000 Belgium. freedom are needed, because, apart from wavelength,
Received 7 October 1999. the BRDF is a function of four variables: two for the
0003-6935兾00兾162592-09$15.00兾0 illumination direction and two for the reflection di-
© 2000 Optical Society of America rection. For isotropic surfaces—those whose BRDF

2592 APPLIED OPTICS 兾 Vol. 39, No. 16 兾 1 June 2000

is invariant under rotation about the surface wavelengths are of interest, and coarse wavelength
normal—the BRDF has just three angular variables, resolution is often sufficient.
requiring three geometric degrees of freedom. In Computer graphics practitioners need complete
this paper we concentrate on this important class of data, but usually lack the resources of an optical
materials. measurement laboratory. It is important to use rel-
We denote the incident and exitant directions by atively nonspecialized equipment in a relatively sim-
共␪i , ␾i 兲 and 共␪e, ␾e兲, with ␪ measuring elevation 共␪ ⫽ 0 ple procedure. Because of these demands,
at the surface normal兲 and ␾ measuring azimuth. researchers in computer graphics have developed
An isotropic BRDF depends only on the difference in techniques and instruments specifically to acquire
azimuth, ⌬␾ ⫽ ␾e ⫺ ␾i , so the variables ␪i , ␪e, and ⌬␾ the data needed.
suffice in the isotropic case. Ward4 used a curved mirror to gather light scat-
In a classical setup,5,6 the three or four angular tered from a flat sample into a CCD camera with a
dimensions are handled by specialized mechanisms fish-eye lens. By use of a semisilvered mirror, the
that position a light source and a detector at various camera captures the entire exitant hemisphere at
directions from a flat sample of the material to be once for each illumination direction, leaving 2 de-
measured. The final dimension, that of wavelength, grees of freedom to handle mechanically. Ward’s in-
is handled either with a broadband spectroradiom- strument is limited by its optics; the hemispherical
eter that measures the entire spectrum at once or by mirror only approximates the ideal ellipsoid used
multiple measurements varying the wavelength of a later by Davis and Rawlings, and vignetting limits
narrow-band source or detector. Because three, the quality of measurements near grazing exitance.
four, or five dimensions must be sampled sequen- Ward used the red-green-blue 共RGB兲 filters of his
tially, the measurement of reflectance functions can CCD camera to yield three bands of spectral infor-
be time-consuming, even with modern computer con- mation.
trols. Moving the motor stages and measuring the Ikeuchi and Sato11 present a system for estimating
reflected light can take several seconds, and, because reflectance model parameters using a surface model
measurements are taken point by point, even a from a range scanner and a single image from a video
sparse sampling of the incident and exitant hemi- camera. In contrast to Ward’s method, they use a
curved sample to capture a set of directions spanning
spheres can take several hours.
a large range of both incident and exitant directions.
The advent of good two-dimensional 共2-D兲 image
Because their goal is to fit a reflectance model, they
sensor arrays, such as CCD’s, offers the opportunity
use a single image and make no attempt to sample
to reduce dimensionality in angle rather than in
the BRDF exhaustively.
wavelength. The sensor array can measure a 2-D
More recently, Lu et al.12 use photographs of a
range of angles simultaneously, leaving one or two
cylindrical sample to give broad angular coverage in
dimensions of angle and one dimension of wavelength the incidence plane, using multiple images with dif-
to be sampled by sequential measurements. ferent source positions to cover all angles. Their
Several such instruments have been developed to goals and methods are similar to ours for the special
measure scattering for optical inspection and metrol- case of incidence-plane measurements. However,
ogy purposes. They typically simplify BRDF mea- we present measurements for the full isotropic
surement by any 共or several兲 of the following BRDF, and we verified the accuracy of our results.
restrictions: covering a subset of the scattering Our method improves on this previous research by
hemisphere, using a single incident direction, or mea- measuring isotropic BRDF’s completely and accu-
suring at a single wavelength. Breault Research’s rately, while retaining simplicity, speed, and sam-
OmniScatR device7 uses fiber optics to collect multi- pling density. Our results can be used not only to
ple points on the hemisphere and direct them to a render images, but also to validate reflectance models
CCD array. It covers a 44° cone of the scattering for particular materials or to investigate BRDF’s that
hemisphere in one measurement. Hatab et al.8 and do not conform to existing models.
McNeil and Wilson9 built a device that uses a hemi- In the following sections we describe the specifics of
spherical screen to capture scattered laser light; this our system, give the results of measuring several
screen is imaged onto a CCD array. Davis and Raw- materials, and demonstrate the accuracy of these re-
lings10 at Boeing have recently patented a device us- sults by comparing them with measurements from a
ing an ellipsoidal mirror to gather a large part of the gonioreflectometer.
scattering hemisphere onto a CCD sensor.
Computer graphics, however, demands both full 3. Method
hemispherical BRDF information and complete cov-
erage of the visible spectrum. Exhaustive measure- A. Measurement Setup and Technique
ment of the entire hemispheres of the BRDF is Our system, pictured in Fig. 1, uses digital photo-
required because image synthesis algorithms must graphs to record light reflected from the surface of a
evaluate the BRDF in arbitrary directions. The vis- test sample of known shape and unknown but uni-
ible wavelength spectrum must be covered for correct form BRDF. It measures the three dimensions of an
color reproduction. On the other hand, accuracy isotropic BRDF by using the curvature of the surface
within a few percent is quite acceptable, only visible to vary two of the three angles and a sequence of

1 June 2000 兾 Vol. 39, No. 16 兾 APPLIED OPTICS 2593

 Fig. 1. Measurement setup.

photographs to vary the third. To do this, we illu-

minate the sample from a sequence of known posi-
tions, using a fixed primary camera to take a
photograph for each position 共Fig. 2兲. The pixels in Fig. 3. Mapping from the surface to the BRDF parameter space in
each of the resulting images give measurements of the incidence plane. If both camera and source were at an infinite
light reflected from various points on the surface; the distance from the sample, ␪e ⫺ ␪i would be constant and the diag-
known geometry of the surface, light source, and cam- onal lines would be straight as shown; in our measurements, they
era allows the incident and exitant directions to be are slightly curved.
computed for each pixel. Each pixel in each image
thus provides a measurement of the BRDF for a par-
ticular configuration of incident and exitant direc-
tions. Depending on the sample geometry, we may To measure the BRDF over the complete hemi-
obtain several independent measurements for each sphere for an isotropic surface, we use a spherical
configuration, which can be compared to verify the sample, which has curvature in both directions. In
measurement or averaged to reduce random errors. this case, the same set of light source positions leads
We can completely measure the BRDF in the plane to measurements on a family of 2-D surfaces of 共␪i , ␪e,
of incidence by using a cylindrical sample. Because ⌬␾兲 triples, which together fill the three-dimensional
the sample’s curvature lies along a single direction, 共3-D兲 parameter space of the isotropic BRDF.
we can arrange for the light source, camera, and sur-
face normal always to lie in a plane. Each image B. Apparatus
will give measurements along a range of 共␪i , ␪e兲 pairs,
with ⌬␾ ⫽ 0 共Fig. 3兲. A set of these images, each Our photographic BRDF measurement technique re-
acquired with the light source in a different position, quires a well-characterized camera, a stable and uni-
produces a family of such curves, which together fill form light source, and a means for measuring their
the entire 2-D parameter space of the incidence-plane positions. Also required are curved samples of ac-
BRDF. We move the light source along a path from curately known shape.
near the camera 共where we measure near retroreflec- Our system comprises the following parts:
tion兲 to opposite the camera 共where we measure
• the primary camera: a cooled 12-bit CCD still
grazing-angle reflection兲.
camera 共Photometrics PXL 1300L兲;
• the light source: a consumer electronic flash
powered by a regulated dc power supply 共Nikon
SB-16 and HP 6030A兲;
• a secondary CCD camera for position measure-
ment 共Kodak DCS 420兲; and
• the samples: metal spheres painted with var-
ious sprayed coatings and metal cylinders painted
similarly or wrapped with sheets of flexible material.

The primary camera, which remains fixed, makes

the actual measurements of radiance reflected from
the test sample. We adjust its overall sensitivity
using neutral-density filters as appropriate to allow
measurement of bright reflections without clipping.
Fig. 2. Schematic of the measurement setup. If a wavelength-dependent measurement is required,

2594 APPLIED OPTICS 兾 Vol. 39, No. 16 兾 1 June 2000

 Fig. 4. Examples of photogrammetric targets.

we make sequential exposures using color separation

filters.
We attach the secondary camera to the light source
to measure the position of the source using automatic
photogrammetry. Each measurement exposure is
made by our opening the shutter of the primary cam-
era, then triggering the secondary camera during the
exposure. The secondary camera triggers the flash, Fig. 5. Field flatness of camera: normalized pixel response ver-
so we obtain a calibration image directly correlated sus radius for image area used.
with the source position, acquired at exactly the same
time as the measurement image. Each calibration
image includes a field of machine-readable targets
nonuniformity by imaging the exit port of an inte-
with known 3-D positions relative to the test sample
grating sphere 共Labsphere CSTM-USS-1200兲. The
and the primary camera. These targets are visible
variation over the entire image plane was of the order
in Fig. 1, below the test sample, and actual-size ex-
of 10%, but we used only the central portion where
amples are shown in Fig. 4. By analyzing these im-
variation was limited to within 2% of uniform, as
ages, we can automatically determine the poses of the
shown in Fig. 5.
secondary camera. Both an a priori error estimate
We modeled our source as a point with uniform
and residuals from the fitting processes involved 共see
angular distribution. To confirm this, we measured
Subsection 3.D兲 confirm that these poses are accurate
the angular variation of the source by photographing
to within a few millimeters. With the light source
a white surface illuminated by the flash; this showed
rigidly attached to the camera, its position is easily
the irradiance to be uniform to within ⫾3% over the
found for each exposure.
⫾5° range of angles we use, as shown by the contour
Because we measure the light source position in
plot in Fig. 6. We reduced the variation from flash to
this way, the light source can be placed freely; the
flash, shown in Fig. 7, to less than 1% from the mean
light source and secondary camera are placed on a
by replacing the flash’s battery pack with a regulated
tripod and moved manually from one position to the
voltage. We also verified that the flash’s output was
next over the course of the measurement. We typi-
essentially unpolarized.
cally use 32 light source positions ranging from 1.3 to
These steps allow radiance and irradiance to be
2.5 m from the sample.
determined in relative units, but we still need to es-
C. Calibration tablish a global scale factor to compute the absolute
BRDF. We measured this factor by illuminating a
To interpret a pixel value as a measurement of
BRDF, the following must be known:

• the responsivity of the pixel sensor to scene ra-

diance;
• the irradiance due to the light source at the
relevant surface point; and
• the geometric arrangement of the surface nor-
mal, the viewing direction, and the illumination di-
rection.

We undertook several calibration steps to ensure that

each of these items was well controlled.
Because our primary camera uses a CCD array
共Kodak KAF-1300i, serial no. MA-3254, chilled to
⫺25 °C兲, we expect that the response of each CCD
element is directly proportional to scene radiance.13,14
This was confirmed experimentally for our camera.15
Physical variation between sensors and limitations of
the lens optics cause a variation in the constant of Fig. 6. Contour plot of relative flash irradiance: contour lines at
proportionality across the image. We measured this ⫾2.5% mean irradiance of center region.

1 June 2000 兾 Vol. 39, No. 16 兾 APPLIED OPTICS 2595

 Fig. 8. Radiance across face of flash 共nonlinear mapping兲.

evenly distributed across its face. Measurements

Fig. 7. Repeatability of flash with 30-s cycle time.
based on an image of the front of the flash 共Fig. 8兲
suggest that it has an effective radiating surface of 50
calibrated diffuse reference sample 共Spectralon SRS- mm ⫻ 13 mm. From the maximum 共2.6-m兲 and min-
99-010, calibrated by Labsphere in Report 14424-A兲 imum 共1.3-m兲 distances used in the measurements,
in a known geometry and measuring the camera sig- we can calculate
nal through each of the color filters. All BRDF mea-
surements were then relative to the known BRDF of 0.050 ⫻ 0.013
⍀min ⫽ ␲
the calibration sample. 2.62
The final prerequisite to BRDF measurement is to
⫽ 3.0 ⫻ 10⫺4 sr, (3)
know the incident and exitant directions. To calcu-
late these, we must know the positions of the light
0.050 ⫻ 0.013
source, test sample, and camera. Any errors in the ⍀max ⫽ ␲
direction to the camera or light source will cause 1.32
errors in the calculated angles corresponding to each ⫽ 1.2 ⫻ 10⫺3 sr. (4)
pixel; also, every 1% error in light source distance
leads to a 2% error in the estimated irradiance. To Our cylindrical samples were aluminum tubing of
establish the positions of the targets used to measure nominally 152 mm outside diameter; the spheres
the location of the light source 共see Subsection 3.B兲, were spun copper, with a diameter of nominally 203
we used the secondary camera and the photogram- mm. The size of each sample was measured in the
metric technique of self-calibrating bundle adjust- image plane, and camera parameters were adjusted
ment.16,17 This determines the targets’ positions to from the size and position of the image in the primary
submillimeter precision, as estimated from residual camera.
errors in the least-squares fit.
We estimate the solid angle of our detector from the D. Data Processing
f-number of the lens, ranging from f:11 to f:22 in The data that result from the measurement of a sin-
these measurements, and the distance from detector gle BRDF are
to sample, which was 1.3 m. From the nominal
28-mm focal length of the lens, we can calculate • 32 measurement images from the primary cam-

冉 冊 2
era 共or 96 when three filters are used for RGB兲, with
0.028 a record of the lens aperture and neutral-density fil-
22 ter used for each; and
⍀min ⫽ ␲ • 32 light source calibration images from the sec-
1.32
ondary camera.
⫽ 3.0 ⫻ 10⫺6 sr, (1)

冉 冊
The major steps involved in processing these data are
2
0.028
11 1. Use the targets visible in the calibration im-
⍀max ⫽ ␲ ages to establish the poses of the secondary camera,
1.32 and from them compute the light source positions.
⫽ 1.2 ⫻ 10⫺5 sr. (2) 2. Find the image of the test sample in one of the
photographs from the primary camera, and use its
The effective solid angle of the source is somewhat size and position to determine the sample’s position.
harder to estimate, both because of the varying dis- 3. Sample the images to obtain measurements
tance from the sample and because its output is not covering part or all of the surface’s BRDF.

2596 APPLIED OPTICS 兾 Vol. 39, No. 16 兾 1 June 2000

 The first step is performed by fully automatic soft-
ware, which locates targets in each secondary camera
image and calculates camera poses. The targets in
each calibration image are used as input to a nonlin-
ear fitting procedure that solves for the camera pose
by minimizing the image-space discrepancy between
the projections of the targets’ known 3-D positions
and their measured locations in the image.
The only human intervention required is in the
second step, where the user clicks on a number of
points along the boundary of the sample in one mea-
surement image; these points are used to allow sub-
pixel precision in the determination of the position
and size of the sample’s image.
The final step works by ray tracing: For each
pixel in the primary camera image, the corresponding Fig. 9. Incidence-plane BRDF measurements plotted against in-
surface point and normal are determined, then the cident and exitant angles.
direction of illumination is computed. The relative
irradiance is computed from the known source geom-
etry; then the relative value of the BRDF is computed
by division of the pixel value, which is proportional to subset of the total BRDF. Even for such a diffuse
radiance, by the irradiance. All that remains is to surface, we see noticeable directional variation; the
normalize this relative value to the measured and graph of an ideal Lambertian surface would be a flat
known BRDF value of our reference sample. This plane. Reciprocity demands that the plot be sym-
computation can be carried out once per pixel to ob- metric left to right, providing an initial check on the
tain a scattered set of raw samples, which will fall on data.
a set of 2-D surfaces, or sheets, in the 3-D domain of We further check our measurements quantita-
the BRDF. tively, both for reciprocity and by comparing against
The generation of BRDF values at arbitrary points an instrument of known accuracy. Reciprocity can
requires the reconstruction of a full 3-D function. be evaluated from Fig. 10, in which the dashed curves
Where such arbitrary evaluation was necessary for show the data for three fixed values of ␪i and the solid
the figures in this paper, we employed a local qua- curves show the data for fixed ␪e. Each pair of
dratic regression procedure,18 using an elliptical ker- curves represents two independent measurements:
nel to span the distances between the sheets defined The data are measured from different parts of the
by individual images without unnecessarily blurring surface, except in the specular direction. Note the
the data along the sheets. close correlation for all angles up to 85°. The rms
deviation over all data out to 75° is 1.5%, and data out
4. Verification to 85° still satisfy reciprocity to within 6%. The er-
To verify the technique, we began with a cylindrical ror figures reported in this paper are the rms average
sample, which allowed measurement of BRDF in the of the relative error between the two sets of data,
incidence plane. This type of measurement pro- after linear interpolation to match the sample pat-
vided simpler data for verification and analysis be- terns. Figure 11 compares the data from the image-
fore we moved on to measuring the full BRDF. It based measurements with independently measured
also allowed us to average several scan lines to reduce data from the gonioreflectometer. The discrepan-
noise. We also measured flat samples of each mate-
rial independently, using a gonioreflectometer that
was designed and verified for accuracy within 5%.19
We validated the image-based measurements both by
verifying reciprocity and by comparing the data from
the two independent measurement systems. The
new method gives results comparable in accuracy to
the gonioreflectometer: Consistency is excellent out
to 75° incidence 共or exitance兲 and reasonable out to
approximately 85°.
We measured a piece of ordinary photocopy paper
共Xerox 4200DP, 20 lb.兲 by wrapping it around a cyl-
inder. The resulting measurements fall on a planar
slice of the BRDF, yielding a function of two angles:
␳bd共␪i , ␪e兲. In Fig. 9, we plot the data as a height field Fig. 10. Reciprocity comparison for white paper. BRDF mea-
over the 共␪i , ␪e兲 plane; each curve contains the data surements plotted for three fixed incident and exitant angles.
from a single image. The dense sampling of our Dashed curves are the fixed incident angles; solid curves are fixed
method yields over 4000 points even for this small exitant angles.

1 June 2000 兾 Vol. 39, No. 16 兾 APPLIED OPTICS 2597

 Fig. 13. Photographs of actual test samples used.

Fig. 11. Gonioreflectometer comparison for white paper. Image-

based BRDF measurements 共solid curves兲 and gonioreflectometer
measurements of the same material 共dashed curves兲.

cies between the two systems are 2.8% 共to 75°兲 and
3.7% 共to 85°兲.
We also measured a flat gray primer; this surface
shows greater deviation from a Lambertian BRDF,
generating a much greater dynamic range. A reci-
procity comparison is shown in Fig. 12; the data are
plotted at two scales to show the entire range of mag-
nitudes. Reciprocity held within 2.5% to 75° and
within 7.8% to 85°.
5. Results
We used the system described above to measure the
full isotropic BRDF of three paints, shown photo-
graphed in Fig. 13. These are a gray primer, a blue
enamel, and a red metallic automotive lacquer. The
Fig. 14. BRDF sampling pattern for a single incident direction.
Each ring is derived from a single digital image.

red and blue paints were coated with a gloss-reducing

finish. For the two colored paints we used RGB color
separation filters, and for the primer we made a sin-
gle broadband measurement. The data for each
color channel of each BRDF comprise some 1.5 mil-
lion samples; approximately 6000 of these, corre-

Fig. 12. Reciprocity comparison for gray primer. BRDF mea-

surements are plotted for three fixed incident and exitant angles.
Dashed curves are fixed incident angles; solid curves are fixed Fig. 15. Scattering diagram for the gray primer at ␪ ⫽ 45°. The
exitant angles. scale is double that used for the red and blue paints.

2598 APPLIED OPTICS 兾 Vol. 39, No. 16 兾 1 June 2000

 with general-purpose equipment. We achieve accu-
racy rivaling that of a specialized gonioreflectometer
but with much greater speed and angular resolution.
The technique is rapid because the two dimensions of
a camera image sample two angular degrees of free-
dom instantaneously, leaving only one to be handled
by sequential measurement. Our system can ac-
quire 1.5 million monochrome samples covering an
entire isotropic BRDF in less time than it takes to
prepare a test sample. The resulting data are inter-
nally consistent and agree closely with independent
measurements.
Fig. 16. Resampled scattering diagrams of the BRDF measure- Our technique is limited to measurement of curved
ments of two paints: a blue enamel 共top row兲 and a red automo- surfaces with uniform BRDF’s. To measure the full
tive lacquer 共bottom row兲. The RGB color measurements are hemispherical BRDF, we must apply the finish to a
shown from left to right. compound-curved surface by spraying, dipping,
grinding, etc., but we can also measure the incidence-
plane reflectance of any material that can be wrapped
around a cylinder.
We have demonstrated measurements of isotropic
BRDF’s, but the technique can also be extended to
measure the BRDF of an anisotropic surface. To do
this, we must rotate the sample to provide a fourth
degree of freedom. Either a spherical sample with
uniform anisotropic BRDF and known orientation
can be used or anisotropic samples of flat, flexible
materials can be measured in the plane of incidence
when the sample is rotated on the cylinder to achieve
different values of ␾.
We also plan to extend our technique to glossier
Fig. 17. Image-based measurements of the red paint 共top row; surfaces. Our detector handles a dynamic range of
identical to bottom row of Fig. 16兲 with the corresponding mea- the order of 3.5 orders of magnitude, so we applied a
surements from the gonioreflectometer 共bottom row兲. The go-
low-gloss finish to the colored paints to simplify this
nioreflectometer data were triangulated directly from the sample
points, whereas the image-based data, which do not come in sets of
initial demonstration of the technique. We can ex-
fixed ␪, were resampled. tend our dynamic range by making multiple expo-
sures through different neutral-density filters and
combining them as did Debevec and Malik.20 As we
measure glossier surfaces, we will also have to ac-
sponding to a single exitant angle, are plotted in Fig. count for the finite solid angles of source and detector.
14. Each ring of points is extracted from a single This method already uses equipment much less
measurement image. expensive than a traditional gonioreflectometer 共the
Figure 15 shows a 3-D scattering diagram of the gonioreflectometer used for validation cost more than
primer for an incident angle of 45°. The surface is four times as much as the equipment used for our
clearly non-Lambertian, showing a broad forward- image-based measurements兲, but the cost of our
scattering peak. Figure 16 shows similar plots for setup could be reduced even further.
the RGB wavelength bands of the two colored paints.
We measured the red paint with the gonioreflecto- • A simpler secondary camera could be used;
meter to validate our results over the entire hemi- color is unneeded, and lower resolution would still
sphere. Figure 17 shows the gonioreflectometer locate the light source with sufficient accuracy. Sev-
measurements alongside the image-based measure- eral consumer CCD camera models should be adapt-
ments from Fig. 16. These plots indicate that our
able for this purpose.
technique has successfully captured the BRDF. The
• A less expensive primary camera could be used,
differences in red and blue channels are accounted
but the higher noise of such a camera would reduce
for, we believe, by the different wavelength ranges of
dynamic range. Multiple-exposure techniques could
the two instruments; the CCD camera’s sensitivity
recover some dynamic range, but would ultimately be
extends beyond the gonioreflectometer’s limit of
limited by blooming in the CCD and optical flare.
⬇710 nm, and our system’s signal-to-noise ratio is far
superior below approximately 450 nm. The authors were supported by the National Sci-
ence Foundation 共NSF兲 Science and Technology Cen-
6. Conclusions and Future Work ter for Computer Graphics and Scientific
We have described a simple technique that can ex- Visualization 共ASC-8920219兲 and by NSF grant ASC-
haustively measure the BRDF of many materials 9523483. S. R. Marschner was also partly sup-

1 June 2000 兾 Vol. 39, No. 16 兾 APPLIED OPTICS 2599

 ported by the Hewlett-Packard Corporation, who also terometer apparatus and process,” U.S. patent 5,241,369 共31
donated several of the workstations used in this re- August 1993兲.
search. Measurement equipment was provided by 10. K. J. Davis and D. C. Rawlings, “Directional reflectometer for
measuring optical bidirectional reflectance,” U.S. patent
NSF grant CTS-9213183 and by a donation from the
5,637,873 共10 June 1997兲.
Imaging Science Division of the Eastman Kodak
11. K. Ikeuchi and K. Sato, “Determining reflectance properties of
Company. an object using range and brightness image,” IEEE Trans.
Pattern Anal. Mach. Intell. 13, 1139 –1153 共1991兲.
References
12. R. Lu, J. J. Koenderink, and A. M. L. Kappers, “Optical prop-
1. F. E. Nicodemus, J. C. Richmond, J. J. Hsia, I. W. Ginsberg,
erties 共BRDF兲 of velvet,” Appl. Opt. 37, 5974 –5984 共1998兲.
and T. Limperis, “Geometric considerations and nomenclature
13. G. C. Holst, CCD Arrays, Cameras, and Displays 共SPIE, Bel-
for reflectance,” Natl. Bur. Stand. 共U.S.兲 Monogr. 160 共U.S.
lingham, Wash., 1996兲.
Department of Commerce, Washington, D.C., 1977兲.
14. S. S.-F. Chen, J. W.-C. Li, K. E. Torrance, and S. N. Pattanaik,
2. J. F. Blinn, “Models of light reflection for computer synthesized
“Preliminary calibration of the photometrics PXL1300L CCD
pictures,” Comput. Graph. 11, 192–198 共1977兲.
camera,” Technical Report PCG-96-1 共Cornell University Pro-
3. K. E. Torrance and E. M. Sparrow, “Theory for off-specular
gram of Computer Graphics, Ithaca, N.Y., 1996兲.
reflection from roughened surfaces,” J. Opt. Soc. Am. 57, 1105–
15. S. N. Pattanaik and K. E. Torrance, “Light measurement using
1114 共1967兲.
4. G. J. Ward, “Measuring and modeling anisotropic reflection,” the photometrics PXL1300L CCD camera,” Technical Report
Comput. Graph. 26, 265–272 共1992兲. PCG-98-1 共Cornell University Program of Computer Graphics,
5. American Society for Testing and Materials, Standard Practice Ithaca, N.Y., 1998兲.
for Angle Resolved Optical Scatter Measurements on Specular 16. C. S. Fraser, M. R. Shortis, and G. Ganci, “Multi-sensor system
or Diffuse Surfaces 共American Society for Testing and Materi- self-calibration,” in Videometrics IV, S. F. El-Hakim, ed., Proc.
als, West Conshohocken, Pa., 1996兲, Standard E 1392–96. SPIE 2598, 2–18 共1995兲.
6. K. E. Torrance and E. M. Sparrow, “Off-specular peaks in the 17. J. H. Chandler and C. J. Padfield, “Automated digital photo-
directional distribution of reflected thermal radiation,” ASME grammetry on a shoestring,” Photogramm. Rec. 15, 545–559
J. Heat Transfer 88, 223–230 共1966兲. 共1996兲.
7. R. J. Castonguay, “New generation high-speed high-resolution 18. J. Fan and I. Gijbels, Local Polynomial Modeling and Its Ap-
hemispherical scatterometer,” in Optical Scattering: Applica- plications 共Chapman & Hall, London, 1996兲.
tions, Measurements, and Theory II, J. C. Stover, ed., Proc. 19. S. C. Foo, “A gonioreflectometer for measuring the bi-
SPIE 1995, 152–165 共1993兲. directional reflectance of material for use in illumination
8. Z. R. Hatab, J. R. McNeil, and S. S. H. Naqvi, “Sixteen-megabit computation,” M.S. thesis 共Cornell University, Ithaca, N.Y.,
dynamic random access memory trench depth characterization 1997兲.
using two-dimensional diffraction analysis,” J. Vac. Sci. Tech- 20. P. E. Debevec and J. Malik, “Recovering high dynamic range
nol. B 13, 174 –181 共1995兲. radiance maps from photographs,” Comp. Graph. Proc., An-
9. J. R. McNeil and S. R. Wilson, “Two-dimensional optical scat- nual Conference Series, 369 –378 共1997兲.

2600 APPLIED OPTICS 兾 Vol. 39, No. 16 兾 1 June 2000

View publication stats