Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

Explaining Homogeneous
Coordinates & Projective Geometry
(/blog/modern-opengl/explaining-
homogenous-coordinates-and-
projective-geometry/)
24 Feb, 2014 — Category: Modern OpenGL Series (/blog/category/modern-opengl/)

Image courtesy of Joachim Baecker (http://commons.wikimedia.org/wiki/File:Zentralprojektion.jpg).

In this article I'm going to explain homogeneous coordinates (a.k.a. 4D coordinates) as simply as I
can. In previous articles, we've used 4D vectors for matrix multiplication, but I've never really defined
what the fourth dimension actually is. Now it's time to take a closer look at projective geometry.

Also, welcome back! It has been a while since my last post. Hopefully I will find some time in the next
couple of months to finish up the Modern OpenGL Series (http://tomdalling.com/blog/category
/modern-opengl/) of articles. The code for article 08 is done, but writing the article will take some time.

Terminology
Most of the time when working with 3D, we are thinking in terms of Euclidean geometry – that is,
coordinates in three-dimensional space ( X , Y , and Z ). However, there are certain situations
where it is useful to think in terms of projective geometry (http://en.wikipedia.org
/wiki/Projective_geometry) instead. Projective geometry has an extra dimension, called W , in
addition to the X , Y , and Z dimensions. This four-dimensional space is called "projective space,"
and coordinates in projective space are called "homogeneous coordinates."

1 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

For the purposes of 3D software, the terms "projective" and "homogeneous" are basically
interchangeable with "4D."

Not Quaternions
Quaternions look a lot like homogeneous coordinates. Both are 4D vectors, commonly depicted as
(X, Y, Z, W) . However, quaternions and homogeneous coordinates are different concepts, with
different uses.

The contents of this article don't apply to quaternions. If I can find the time, I might write a quaternion
article in the future.

An Analogy In 2D
First, let's look at how projective geometry works in 2D, before we move on to 3D.

Imagine a projector that is projecting a 2D image onto a screen. It's easy to identify the X and Y
dimensions of the projected image:

Now, if you step back from the 2D image and look at the The W dimension is the
projector and the screen, you can see the W dimension
distance from the projector to
too. The W dimension is the distance from the projector to
the screen.
the screen.

2 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

So what does the W dimension do, exactly? Imagine what The value of W affects the
would happen to the 2D image if you increased or decreased
size (a.k.a. scale) of the
W – that is, if you increased or decreased the distance
between the projector and the screen. If you move the
image.
projector closer to the screen, the whole 2D image becomes
smaller. If you move the projector away from the screen, the 2D image becomes larger. As you can
see, the value of W affects the size (a.k.a. scale) of the image.

Applying It To 3D
There is no such thing as a 3D projector (yet), so it's harder to imagine projective geometry in 3D, but
the W value works exactly the same as it does in 2D. When W increases, the coordinate expands

3 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

(scales up). When W decreases, the coordinate shrinks (scales down). The W is basically a
scaling transformation for the 3D coordinate.

When W = 1
The usual advice for 3D programming beginners is to always set W = 1 whenever converting a 3D
coordinate to a 4D coordinate. The reason for this is that when you scale a coordinate by 1 it
doesn't shrink or grow, it just stays the same size. So, when W = 1 it has no effect on the X , Y
or Z values.

For this reason, when it comes to 3D computer graphics, coordinates are said to be "correct" only
when W = 1 . If you rendered coordinates with W > 1 then everything would look too small, and
with W < 1 everything would look too big. If you tried to render with W = 0 your program would
crash when it attempted to divide by zero. With W < 0 everything would flip upside-down and back-
to-front.

Mathematically speaking, there is no such thing as an "incorrect" homogeneous coordinate. Using

coordinates with W = 1 is just a useful convention for 3D computer graphics.

The Math
Now, let's look at some actual numbers, to see how the math works.

3 meters away from the screen, and there is a dot on the 2D image at
Let's say that the projector is
the coordinate (15, 21) . This gives us the projective coordinate vector (X, Y, W) = (15, 21, 3) .

Now, imagine that the projector was pushed closer to the screen so that the distance was 1 meter.
The closer the project gets to the screen, the smaller the image becomes. The projector has moved

4 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

three times closer, so the image becomes three times smaller. If we take the original coordinate
vector and divide all the values by three, we get the new vector where W =1:

15 21 3
( , , ) = (5, 7, 1)
3 3 3

The dot is now at coordinate (5, 7) .

This is how an "incorrect" homogeneous coordinate is converted to a "correct" coordinate: divide all
the values by W . The process is exactly the same for 2D and 3D coordinates.
Dividing all the values in a vector is done by scalar multiplication with the reciprocal of the divisor.
Here is a 4D example:

1 10 20 30 5
(10, 20, 30, 5) = ( , , , ) = (2, 4, 6, 1)
5 5 5 5 5

Written in C++ using GLM, The example above would look like this:

glm::vec4 coordinate(10, 20, 30, 5);

glm::vec4 correctCoordinate = (1.0/coordinate.w) * coordinate;
//now, correctCoordinate == (2,4,6,1)

Uses Of Homogeneous Coordinates In

Computer Graphics

5 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

As mentioned earlier, in regard to 3D computer graphics, homogeneous coordinates are useful in

certain situations. We will look at some of those situations here.

Translation Matrices For 3D Coordinates

Rotation and scaling transformation matrices only require A four-column matrix can only
three columns. But, in order to do translation, the matrices
be multiplied with a four-
need to have at least four columns. This is why
transformations are often 4x4 matrices. However, a matrix
element vector, which is why
with four columns can not be multiplied with a 3D vector, due we often use homogeneous
to the rules of matrix multiplication. A four-column matrix can 4D vectors instead of 3D
only be multiplied with a four-element vector, which is why
vectors.
we often use homogeneous 4D vectors instead of 3D
vectors.

The 4th dimensionW is usually unchanged, when using homogeneous coordinates in matrix
transformation. W is set to 1 when converting a 3D coordinate into 4D, and it is usually still 1
after the transformation matrices are applied, at which point it can be converted back into a 3D
coordinate by ignoring the W . This is true for all translation, rotation, and scaling transformations,
which are by far the most common types of transformations. The notable exception is projection
matrices, which do affect the W dimension.

Perspective Transformation
In 3D, "perspective" is the phenomenon where an object appears smaller the further away it is from
the camera. A far-away mountain can appear to be smaller than a cat, if the cat is close enough to
the camera.

6 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

Perspective is implemented in 3D computer graphics by Perspective is implemented in

using a transformation matrix that changes the W element
3D computer graphics by
of each vertex. After the the camera matrix is applied to each
vertex, but before the projection matrix is applied, the Z using a transformation matrix
element of each vertex represents the distance away from that changes the W element
the camera. Therefore, the larger Z is, the more the vertex of each vertex.
should be scaled down. The W dimension affects the
scale, so the projection matrix just changes the W value
based on the Z value. Here is an example of a perspective projection matrix being applied to a
homogeneous coordinate:

⎡1 0 0 0⎤⎡2⎤ ⎡2⎤
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 1 0 0⎥⎢3⎥ ⎢3⎥
⎢0 0 ⎥⎥ ⎢⎢ 4 ⎥⎥ ⎢⎢ 4 ⎥⎥
=
⎢ 0 1
⎣0 0 1 0⎦⎣1⎦ ⎣4⎦

Notice how the W value is changed to 4 , which comes from the Z value.
After the perspective projection matrix is applied, each Perspective division is just a
vertex undergoes "perspective division." Perspective division
specific term for converting
is just a specific term for converting the homogeneous
coordinate back to W = 1 , as explained earlier in the the homogeneous coordinate
article. Continuing with the example above, the perspective back to W=1
division step would look like this:

1
(2, 3, 4, 4) = (0.5, 0.75, 1, 1)
4

After perspective division, the W value is discarded, and we are left with a 3D coordinate that has
been correctly scaled according to a 3D perspective projection.

In GLM, this perspective projection matrix can be created using the glm::perspective or
glm::frustum functions. In old-style OpenGL, it is commonly created with the gluPerspective or
gluFrustum functions. In OpenGL, perspective division happens automatically after the vertex
shader runs on each vertex. This is one reason why gl_Position , the main output of the vertex
shader, is a 4D vector, not a 3D vector.

Positioning Directional Lights

One property of homogeneous coordinates is that they allow you to have points at infinity (infinite

7 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

length vectors), which is not possible with 3D coordinates. Points at infinity occur when W = 0 . If
you try and convert a W = 0 homogeneous coordinate into a normal W = 1 coordinate, it results
in a bunch of divide-by-zero operations:

1 2 3 4 0
(2, 3, 4, 0) = ( , , , )
0 0 0 0 0

This means that homogeneous coordinates with W = 0 can not be converted into 3D coordinates.
What use does this have? Well, directional lights can be though of as point lights that are infinitely far
away. When a point light is infinitely far away the rays of light become parallel, and all of the light
travels in a single direction. This is basically the definition of a directional light.

So traditionally, in 3D graphics, directional lights are If W = 1 , then it is a point

differentiated from point lights by the value of W in the
light. If W = 0 , then it is a
position vector of the light. If W = 1 , then it is a point light.
If W = 0 , then it is a directional light. directional light.

This is more of a traditional convention, rather than a useful

way to write lighting code. Directional lights and point lights are usually implemented with separate
code, because they behave differently. A typical lighting shader might look like this:

if(lightPosition.w == 0.0){
//directional light code here
} else {
//point light code here
}

Summary
Homogeneous coordinates have an extra dimension called W , which scales the X , Y , and Z
dimensions. Matrices for translation and perspective projection transformations can only be applied to
homogeneous coordinates, which is why they are so common in 3D computer graphics. The X, Y,
and Z values are said to be "correct" when W = 1 . Any homogeneous coordinate can be
converted to have W = 1 by dividing all four dimensions by the W value, except if W = 0 .
When W = 0 , the coordinate represents a point at infinity (a vector with infinite length), and this is
often used to denote the direction of directional lights.

Enjoy this post?

Subscribe via RSS (/feed/) Follow @tom_dalling (https://twitter.com/tom_dalling)

8 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

Comments

9 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

40 Comments Tom Dalling  Login

Sort by Best
 Recommend 19 ⤤ Share

Join the discussion…

LOG IN WITH
OR SIGN UP WITH DISQUS ?

Name

Ian Vincent • 4 years ago

This is the best explanation that I've seen so far.
3△ ▽ • Reply • Share ›

Shoubin Zhao • 4 years ago

awesome!
3△ ▽ • Reply • Share ›

Zeki Zeybek • 4 months ago

Sağolun hocam sayenizde guvenlık gorevlısı oldum
1△ ▽ • Reply • Share ›

Shruti Manvikar • a year ago

Wonderful article! Never seen it explained better than this.
1△ ▽ • Reply • Share ›

ShuddaSky • 2 years ago

Tom's tutorials sure do point you in the right direction.
1△ ▽ • Reply • Share ›

Sundaram • 3 years ago

Great article, and the pictures would get makes it easier to get the point across! A couple of points:

A. The function name is glFrustum (and not gluFrustum) which was part of the official OpenGL
specification 2.1 and was marked deprecated in 3.0 and removed in 3.1

B. Focal length is directly proportional to zoom; increasing it makes the projection of objects larger,
while FoV is inversely proportional to focal length and thereby to zoom. Thus

with W>1 then everything would look too small, and with W<1 everything would look too big.

should perhaps be reversed. Think about it. A candle would project a ball on the wall behind it. Now
moving the wall closer to the object (or the other way) would make the shadow smaller and make
them farther would make the shadow larger.
1△ ▽ • Reply • Share ›

Kang Weon Jeong • 4 years ago

Thanks a lot, you're so gentle.
1△ ▽ • Reply • Share ›

caiyue Li • a month ago

10 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

Subscribe

 RSS (/blog/feed/)

 Twitter (https://twitter.com/tom_dalling)

Recent Posts

The Pure Function As An Object (PFAAO) Ruby Pattern (/blog/ruby/pure-function-as-an-object-

PFAAO-pattern/)

FizzBuzz In Too Much Detail (/blog/software-design/fizzbuzz-in-too-much-detail/)

Making Fruity Bat (a Flappy Bird clone) in Ruby (/blog/ruby/fruity-bat-flappy-bird-clone-in-ruby/)

Modern OpenGL 08 – Even More Lighting: Directional Lights, Spotlights, & Multiple Lights (/blog
/modern-opengl/08-even-more-lighting-directional-lights-spotlights-multiple-lights/)

OpenGL in 2014 (/blog/modern-opengl/opengl-in-2014/)

Blog Categories

Cocoa (/blog/category/cocoa/) ( 5  (/blog/category/cocoa/feed/))

Coding Style/Conventions (/blog/category/coding-styleconventions/) ( 3  (/blog/category/coding-

styleconventions/feed/))

Coding Tips (/blog/category/coding-tips/) ( 4  (/blog/category/coding-tips/feed/))

Miscellaneous (/blog/category/random-stuff/) ( 1  (/blog/category/random-stuff/feed/))

Modern OpenGL Series (/blog/category/modern-opengl/) ( 10  (/blog/category/modern-opengl

/feed/))

Ruby (/blog/category/ruby/) ( 2  (/blog/category/ruby/feed/))

Software Design (/blog/category/software-design/) ( 9  (/blog/category/software-design/feed/))

Software Processes (/blog/category/software-processes/) ( 1  (/blog/category/software-

processes/feed/))

Web (/blog/category/web/) ( 1  (/blog/category/web/feed/))

Blog Archives

11 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

February 2016 (/blog/2016/02/) ( 1 )

April 2015 (/blog/2015/04/) ( 1 )

February 2015 (/blog/2015/02/) ( 1 )

November 2014 (/blog/2014/11/) ( 1 )

September 2014 (/blog/2014/09/) ( 1 )

February 2014 (/blog/2014/02/) ( 1 )

April 2013 (/blog/2013/04/) ( 1 )

March 2013 (/blog/2013/03/) ( 1 )

February 2013 (/blog/2013/02/) ( 1 )

January 2013 (/blog/2013/01/) ( 2 )

December 2012 (/blog/2012/12/) ( 2 )

November 2012 (/blog/2012/11/) ( 2 )

July 2012 (/blog/2012/07/) ( 2 )

May 2012 (/blog/2012/05/) ( 2 )

March 2012 (/blog/2012/03/) ( 2 )

December 2011 (/blog/2011/12/) ( 1 )

December 2010 (/blog/2010/12/) ( 1 )

November 2010 (/blog/2010/11/) ( 1 )

May 2010 (/blog/2010/05/) ( 1 )

April 2010 (/blog/2010/04/) ( 1 )

February 2010 (/blog/2010/02/) ( 1 )

December 2009 (/blog/2009/12/) ( 1 )

November 2009 (/blog/2009/11/) ( 3 )

October 2009 (/blog/2009/10/) ( 2 )

12 of 13 4/21/18, 4:18 AM
Explaining Homogeneous Coordinates & Project... https://www.tomdalling.com/blog/modern-openg...

July 2009 (/blog/2009/07/) ( 1 )

June 2009 (/blog/2009/06/) ( 1 )

May 2009 (/blog/2009/05/) ( 1 )

© 2009 – 2016 Tom Dalling

 (https://twitter.com/tom_dalling)  (https://github.com/tomdalling)  (https://stackoverflow.com/users/108105/tom-

dalling)  (mailto:tom at tomdalling com)

13 of 13 4/21/18, 4:18 AM