Homogeneous Coordinates

Jules Bloomenthal and Jon Rokne

Department of Computer Science

The University of Calgary

Introduction

Homogeneous coordinates have a natural application to Computer Graphics; they form a basis for

the projective geometry used extensively to project a three-dimensional scene onto a two-

dimensional image plane. They also unify the treatment of common graphical transformations and

operations. The graphical use of homogeneous coordinates is due to [Roberts, 1965], and an early

review is presented by [Ahuja, 1968]. Today, homogeneous coordinates are presented in numerous

computer graphics texts (such as [Foley, Newman, Rogers, Qiulin and Davies]); [Newman], in

particular, provides an appendix of homogeneous techniques. [Riesenfeld] provides an excellent

introduction to homogeneous coordinates and their algebraic, geometric and topological

significance to Computer Graphics. [Bez] further discusses their algebraic and topological

properties, and [Blinn77, Blinn78] develop additional applications for Computer Graphics.

Homogeneous coordinates are also used in the related areas of CAD/CAM [Zeid], robotics

[McKerrow], surface modeling [Farin], and computational projective geometry [Kanatani]. They

can also extend the number range for fixed point arithmetic [Rogers].

Our aim here is to provide an intuitive yet theoretically based discussion that assembles the key

features of homogeneous coordinates and their applications to Computer Graphics. These

applications include affine transformations, perspective projection, line intersections, clipping, and

rational curves and surfaces. For the sake of clarity in accompanying illustrations, we confine our

1
 2

development to two dimensions and then use the intuition gained to present the application of

homogeneous coordinates to three dimensions. None of the material presented here is new; rather,

we have tried to collect in one place diverse but related methods.

[Kline] provides a brief history of homogeneous coordinates, crediting Mbius with their

introduction [Mbius]. Given a fixed triangle in the plane, Mbius defined a set of homogeneous

coordinates for a point p to be the weights required at the triangle vertices such that p became the

center of gravity of the triangle (Figure 1, left). The point p is computed as (waa, wbb, wcc), with

the condition that wa+wb+wc = 1. The three unknowns wa, wb, and wc, are called the barycentric

coordinates of p with respect to a, b, and c. [Snyder] provides an algorithm for computing wi;

[Farin] relates w to area by wa = Aa/(Aa+Ab+Ac) and similarly for wb and wc (Figure 1, middle).

Plcker defined another set of homogeneous coordinates by considering the signed distances from

a point to the edges of a fixed triangle (Figure 1, right); here, p = (la, lb, lc).

b b b
wb Ac lc
a a Aa a la
balance Ab lb
wa point p
c c c
wc

Figure 1: barycentric and Plcker coordinates.

Barycentric and Plcker coordinates are examples of coordinate systems in which n+1 values

represent an n-dimensional point. One immediate attribute of these systems is the invariance of a

point when scaled: scaling the Mbius weights or the size of the Plcker triangle does not change

the position of p. As an aside, we note that higher dimensional Plcker coordinates have been

applied to ray-tracing [Pellegrini] (for an exposition of n-dimensional Plcker coordinates, also

known as Grassmann coordinates, see [Stolfi]).

Homogeneous Coordinates for the Projective Plane

 3

Plcker realized that the homogeneous coordinates [x, y, w] provided a scale invariant

representation for points (x, y) in the Euclidean plane, with x ~ x/w, y ~ y/w, and w 0. He

regarded homogeneous points with w = 0 as corresponding to points in the ordinary plane because

they are infinitely far away. The need for such points at infinity arose from the 16th and 17th

century work of Kepler and Desargues; both realized that a parabola has two foci, one finite and

one infinite [Coxeter].

The ordinary plane augmented with points at infinity is known as the projective plane.

[Aleksandrov] describes the projective plane by considering all lines and planes passing through a

given point s; if they are intersected by a plane P that does not pass through s, then each point (or

line) on P may be associated with a line (or plane) through s, as shown in Figure 2, left. This does

not quite imply a one-to-one mapping of lines (or planes) through s with points (or lines) on P,

because those lines or planes through s parallel to P do not intersect P. By convention, however,

the parallel lines are said to intersect P at ideal (infinitely distant) points; the parallel plane is said

to intersect P at the ideal line. Thus, the projective plane, which we will call 2, is considered to

contain both ideal points and the ideal line. The projective plane cannot be represented within the

finite Euclidean coordinate system; it can, however, be represented by homogeneous coordinates,

which is the fundamental reason for their use in projective geometry.

w
V
piercing w=1
w=1
P piercing S2

lines on 2
x
s center of projection 3
S2

Figure 2: the projective plane.

Projective geometry is, in a sense, the geometry of imaging. This was already understood, for

example, by Albrecht Drer who used a mechanical device to draw objects in perspective [Penna].
 4

It is, therefore, a natural tool for Computer Graphics, as discussed in [Herman, Penna]. General

projective geometry is discussed in several texts [Ayres, Coxeter, Ryan] and [Veblen] provides an

excellent development of homogeneous coordinates and projective and metric geometry,

beginning with an elementary set of assumptions.

Mathematically, the mapping from planes and lines through s to lines and points on the projective

plane is the transformation of the usual Euclidean space into projective space. The following

statements each define the two-dimensional projective space, 2 (from [Ryan]):

1) The set of all equivalence classes of ordered triples of non-zero vectors in 3,

where equivalence is the mutual proportionality of two vectors.

2) The set of all lines passing through the origin of 3.

3) The set of all pairs of antipodal points of S2, the unit sphere in 3.

We denote the usual two-dimensional Euclidean space (also known as physical, ordinary, or

proper space) by 2. Within 2, each point is represented as a two-component vector, (x, y),

where both x and y are finite values in an orthogonal coordinate system. The relationships

between 3, the unit sphere, S2, and the projective plane, 2, are illustrated in Figure 2, right.
Points in 2 are lines through the origin, as in the second definition above. This means that

equivalence classes of coordinates of 3 represent points in 2.

In practice, homogeneous coordinates represent 2 by mapping each Euclidean point (x, y)

2 to [x, y, w] 3 (w 0), which is a member of the equivalence class of points in 2. The

mapping is achieved by the equivalences x ~ x/w and y ~ y/w (we enclose Euclidean

coordinates within parentheses and homogeneous coordinates within brackets; their equivalence is

signified by ~).

2 is not a vector space in the same manner as 2. Indeed, scalar multiplication simply forms a
 5

new representative of the equivalence class. Furthermore, if the representatives are chosen from

the intersection with the plane w = 1, then the vector addition of two points in 2 computes a

representative of the midpoint between the two vectors (for example, given two ordinary vectors,

(1, 2) and (3, 4), addition of their homogeneous equivalents is [1, 2, 1]+[3, 4, 1] = [4, 6, 2] ~ (2, 3)).

The division by w means that the conversion of a homogeneous point to its Euclidean equivalent is

inherently a projection of the homogenous point onto the w = 1 plane. Figure 2, right, illustrates

this projection in the two-dimensional case; three-dimensional homogeneous points on the S2

sphere are projected onto the w = 1 plane. [Riesenfeld] provides an illustration of four-

dimensional homogeneous points projected onto a three-dimensional hyper-plane.

Because homogeneous points represent projection, they can also represent points at infinity.

Consider a homogeneous point as w approaches 0; for example, in Figure 3, [2, 3, w] is shown for w

= {4, 2, 1, 1/2, 0}. As w approaches 0, the projected Euclidean points move away from the origin

in the (2, 3) direction. At w = 0, the point is infinitely far and may be treated as a positionless

vector.

(2, 3, 4)
w
w = 1 plane line from (0, 0, 1)
(2, 3, 2) towards (2, 3, 1)
(2, 3, 1)
8

y (2, 3, .5)
(2, 3, 0)
homogeneous point
(0, 0, 0) x
projection to w = 1 plane

Figure 3: projection to the w = 1 plane.

We close this section with a brief reflection on the term homogeneous, which [Oxford] defines as:

1) Of the same kind so as to be commensurable.

2) Of the same degree or dimension: consisting of terms of the same dimension.

 6

With this in mind, consider a general conic in 2:

f(x, y) = ax2 +by2 +cxy +dx +ey + f = 0.

The homogeneous form for the conic is found by replacing x with x/w and y with y/w;

multiplying by w2 yields:

f*(x, y, w) = ax2+by2 +cxy +dxw +eyw +fw2 = 0.

Here, all terms have total variable degree two, as in definition 2), above. In passing, we note that

any polynomial function in 2 has a form equivalent to f and may be transformed, via a change to

homogeneous coordinates, to a polynomial function with constant total degree for each term. This

holds for any polynomial function in a finitely dimensioned space.

Homogeneous Coordinates for Two Dimensions

An important, practical aspect of the homogeneous coordinate system is its unification of the

translation, scaling and rotation of geometric objects. In 2 Euclidean space, the most general

affine mapping is

p = pA+c, or (px, py) = (px, py)A+(cx, cy),

where p and p are points in 2, A is a 2 by 2 matrix representing scaling and rotation, and the

point c represents translation. In this formulation, translation is treated differently from rotation

or scaling.

If, however, p is represented as a homogeneous point [px py 1] and a 2 3 matrix is employed,

 7

then translation may be treated the same as rotation and scaling:

[ ]
a00 a01

[px py] = [px py 1] a10 a11 .

c1 c2

A more consistent, and ultimately simpler result is obtained by introducing a third column to the

matrix and a third (homogeneous) coordinate to the result:

[ ]
a00 a01 0

[px py 1] = [px py 1] a10 a11 0 .

c1 c2 1

px py are unchanged, but the result is now three-dimensional and the matrix is square. Thus:

the resulting (homogeneous) vector has the same dimension as the argument

(homogeneous) vector,

the matrix is invertible (assuming a non-zero determinant),

two or more transformations may be concatenated.

The above transformation may be represented compactly by bordering the matrix A with the

column vector (0, 1)T and the row vector c:

[p 1] = [p 1]
[ ] A

c
0

1
.

Here [px py 1] and [px py 1] are replaced with their compact forms [p 1] and [p 1] so that the

components conform in size with the sub-matrices. In this formulation, all affine transformations

are matrix multiplications. Computer graphic transformations are usually encoded as matrices

because of their representational simplicity and computational efficiency. We now summarize the
 8

use of homogeneous coordinates for affine transformations.

Affine Transformations

Translation. In the Euclidean coordinate system, translation is p = p+c = pI+c, where I is

the identity matrix. In the homogeneous coordinate system this is

[ ]
1 0 0

[px py 1] = [px py 1] 0

cx cy
1 0 , compactly expressed as [p 1] = [p 1]

1
[ ] I

c
0

1
.

Rotation. In the Euclidean coordinate system, rotation is defined by p = pR,

where R =
[ cos()

sin()
-sin()

cos()
] for a rotation of radians counter-clockwise about the origin.

Expressed with homogeneous coordinates, this becomes

[ ]
cos() -sin() 0

[px py 1] = [px py 1] sin()

0
cos()

0
0 , compactly: [p 1] = [p 1]

1
[ ]
R

0
0

1
.
 9

Scaling. A non-uniform scaling is defined by S =

[ ]
sx

0
0

sy
.

In a homogeneous coordinate system this becomes

[ ]
sx 0 0

[px py 1] = [px py 1] 0

0
sy

0
0 , compactly: [p 1] = [p 1]

1
[ ]
S

0
0

1
.

Sequences

A sequence, or concatenation, of affine transformations can be expressed as a single matrix. For

example, a translation followed by a rotation and then a scaling is equivalent to

[p 1] = [p 1]
[ ][ ][ ] [
I

c
0

1
R

0
0

1
S

0
0

1
= [p 1]
RS

cRS
0

1
]
.

Reducing a sequence of transformations to a single matrix improves computational performance,

especially when numerous points are to be transformed by the same transformation.

Homogeneous Lines

A line l in 2 is defined by ax+by+c = 0, with a, b, c constant; the intersection of lines l1 and l2 is

found by solving
 10

a1x+b1y+c1 = 0

a2x+b2y+c2 = 0,

obtaining: x =
| || | | || |
-c1

-c2
b1

b2
/
a1

a2
b1

b2
; y =
a1

a2
-c1

-c2
/
a1

a2
b1

b2
.

The two lines l1 and l2 intersect in 2 provided they are not parallel. This special provision may be

eliminated by representing a line homogeneously [Newman]:

ax+by+cw = 0 (with w = 1, in general),

and representing the intersection as a homogeneous point:

[x, y, w] =
(| | | | | | )
-c1

-c2
b1

b2
,
a1

a2
-c1

-c2
,
a1

a2
b1

b2
.

Dividing by w yields a finite point in 2 if w 0, and a point at infinity if w = 0 (i.e, the lines are
parallel). [x, y, w] is simply the cross product, [a1, b1, c1] [a2, b2, c2].

If the intersection of two lines is the cross product of their coefficients, the intersection of two

parallel lines may be given as [a, b, c] [a, b, d] = [b, -a, 0], or [1, -a/b, 0], which is the point at

infinity in the direction of the lines slope. Note that the line at infinity may be represented as [0, 0,

a]; the dot product with this line and any point at infinity is zero; thus, all points at infinity lie on

the line at infinity.

We also note the duality between line and point in that the cross product of two homogeneous
 11

points yields the coordinates of their connecting line. This duality between point and line in two

dimensions has, in three dimensions, a corresponding duality between point and plane. See

[Blinn77] for the use of homogeneous coordinates to represent lines in 3.

Conics

In our introduction we briefly touched upon the general second degree implicit curve in 2,
showing that it could be converted to homogeneous form in 2 (represented by points in 3). This

curve is either null, a conic section, a single point, intersecting lines, or coincident lines [Blinn84,

Brieskorn].

It can also be shown that conic sections are equivalent under perspective transformation. As an

example, consider the transformation of a circle to an ellipse, both conic sections. Let such a cone

be defined and let a plane A intersect the cone to form an ellipse. Now intersect the cone with a

plane B normal to the cone axis at a distance from the cone vertex such that the circle of

intersection has the desired radius. The ellipse in plane A is simply the projection of the circle in

plane B, with the axis of the cone being the direction of projection.

A B

Figure 4: conic section as a projection.

We can define second degree curves as the set of all points [x, y, w] that satisfy the general

homogeneous equation of degree 2:

ax2+by2+cxy+dxw+eyw+fw2 = 0, which can be written in matrix form as:

 12

[] [ ]
x a c/2 d/2

[x y w] Q y = 0, where Q = c/2 b e/2 is symmetric.

w d/2 e/2 f

The matrix Q is known as the matrix of the second degree curve. In particular, if the eigenvalues

of Q have the signs (+, +, -) or (-, -, +), then Q represents a conic section (the eigenvalues are

real because Q is symmetric) [Blinn84].

Now, letting T be a scaling transformation, TQTT is the quadratic form for the transformed curve.

We represent the unit circle by

[ ]
1 0 0
Q= 0 1 0

0 0 -1

so that [x, y, w] Q [x, y, w]T = x2+y2-w2 = 0. The matrix for the scaled unit circle is

[ ][ ][ ] [ ]
sx 0 0 1 0 0 sx 0 0 sx2 0 0

0 sy 0 0 1 0 0 sy 0 = 0 sy2 0

0 0 1 0 0 -1 0 0 1 0 0 -1

and the homogeneous second degree curve is x2sx2+y2sy2-w2 = 0, which can be reformulated as
 13

x2 y2 w2
+ - = 0,
sy2 sx2 s2x2y2

which is the homogeneous form for the ellipse [Patterson].

Rational Curves

In many applications the implicit formulation for a curve is declined in favor of a parametric

formulation that expresses the curve in terms of easily manipulated control points. Specifically, the

parametric curve may be defined by

n
C(t) =

i=1
Bi(t)Pi,

where Pi is a set of n control points, Bi is a corresponding set of basis functions, each of degree n-1,

and the parameter t is usually evaluated over (0, 1).

Unfortunately, parametric curves are a limited subset of implicitly defined curves. For example,

second degree polynomial basis functions yield parabolas; no choice of Bi will yield a portion of a

circle, ellipse, or hyperbola [Patterson]. As we have seen, however, all conic sections are equivalent

under projection, and this motivates the use of projection to produce an extended family of

parametric curves. These curves are known as rational curves and may be defined by

n n
C(t) =
wiBi(t)Pi / wiBi(t),
i=1 i=1
 14

where wi are weights applied to the control vertices. Increasing the value of a weight draws the

curve closer to the corresponding control point. Usually wi > 0 (otherwise the resulting curve may

be undefined or contain asymptotes [Patterson]) and usually the basis functions form a partition of

unity, meaning Bi(t) = 1, which implies that the shape of the curve is unaffected by translation
of the control points [Barsky].

The second degree rational parametric curve may be represented in matrix form as

[ ][ ][ ] [ ][ ]
B0(t) T w0 0 0 Px0 Py0 1 B0(t) T w0Px0 w0Py0 w0

B1(t) 0 w1 0 Px1 Py1 1 = B1(t) w1Px1 w1Py1 w1 .

B2(t) 0 0 w2 Px2 Py2 1 B2(t) w2Px2 w2Py2 w2

That is, evaluation of the curve is facilitated by representing the control points in their

homogeneous form. For wi = 1, the above represents a second degree non-rational parametric

curve.

Applications in Three Dimensions

We now apply homogeneous coordinates to three-dimensional Euclidean points. Early in the

development of Computer Graphics, L. G. Roberts noted the value of homogeneous coordinates,

stating, the use of homogeneous coordinates throughout is extremely important in order to

maintain the simplicity of the results, although its original purpose was to allow perspective

transformations [Roberts].
 15

In other words, the use of the additional, homogeneous coordinate not only produces polynomials

of fixed degree, it also provides a method for consistent manipulation of the Euclidean space.

It is well known that a 3 by 3 matrix can represent three-dimensional scaling and rotation, but not

translation. As in the two-dimensional case, translation becomes possible with the addition of a

row and column to the matrix:

[ ]
1 0 0 0

0 1 0 0
[x y z w] = [x y z 1]
0 0 1 0

tx ty tz 1

Although it is possible to represent this in compact form:

[p 1] = [p 1]
[ ]
I

t
0

1
, where I is the identity matrix and t the translation,

we do not do so when considering three-dimensional applications because individual components

of the transformation matrix will be discussed in detail.

Perspective Projection

Fundamental to three-dimensional Computer Graphics is the projection of three-dimensional

objects onto a two-dimensional image plane. This projection is usually a perspective projection

with the center of projection being the point of view (eyepoint) and the central projector

perpendicularly intersecting the projective (image) plane.

In creating shaded images, a perspective projection is not strictly necessary; ray tracing, for

example, computes pixel values directly without a perspective transformation of the object. For

line drawings and polygon rendering, however, the perspective transformation is essential.
 16

[Carlbom] provides a review of planar projections.

As shown in Figure 5, we identify the screen of an output device with a rectangular domain in 2;
the rendering of objects in 3 is naturally performed via their projection to 2.

Eye
Object
Screen

Figure 5: perspective projection onto the raster-screen.

Referring to the side view in Figure 6, the image point (x, y) is computed from the object point

(x, y, z) via similar triangles:

y/y = D/z and x/x = D/z, or (x, y) = (xD/z, yD/z).

(x, y, D)
(x, y, z) Sy
y
Z
(0, 0, 0) D
image window

Figure 6: the perspective projection

The division by z is known as the perspective divide. This non-linear relationship in three

dimensions can be formulated as a linear relationship in four dimensions through the use of

homogeneous coordinates; the foundation for such a formulation is provided by [Bez].

 17

Until now we have used the homogeneous coordinate to accommodate the larger matrix required

to perform affine transformations; after the transformation, the fourth coordinate, w, has remained

1. If, however, the resulting homogeneous coordinate w is proportional to the distance from the

eye to a point, a perspective projection of that point onto the image plane is effected.

For example, let us assume the eye is located at the origin and directed towards the positive z-axis,

with the x-axis to the right and the y-axis up (a left-handed coordinate system); alternatively, as

discussed in [Blinn93], the image plane may be placed at the origin. We can project points onto the

z = D plane with the following matrix transformation:

[ ]
1 0 0 0

0 1 0 0
(x, y, z) = (xD/z, yD/z, D) ~ [x y z z/D] = [x y z 1] .
0 0 1 1/D

0 0 0 0

As D approaches infinity, w = 1/D approaches 0 and the transformed points become infinitely

far; their position parallel to the xy plane is unaffected by z and, effectively, an orthographic

transformation results.

Perspective, Projection, and Perspective Projection

The perspective projection above yields a constant z = D, which is a loss of depth information

due to the linear dependence of the third and fourth columns of the matrix. Introducing a second

non-zero term, e.g. -1, into the third column does not affect x and y, but z becomes D-D/z.

The purpose of this additional term is to compress the Euclidean space z [1, ] to z [0, D].

[ ]
1 0 0 0

0 1 0 0
(x, y, z) = (xD/z, yD/z, D-D/z) ~ [x y z-1 z/D] = [x y z 1] .
0 0 1 1/D

0 0 -1 0
This, then, is the perspective transformation; multiplied by a pure projection it yields the

perspective-projection transformation:
 18

[ ][ ] [ ]
1 0 0 0 1 0 0 0 1 0 0 0

0 1 0 0 0 1 0 0 0 1 0 0
= .
0 0 1 1/D 0 0 0 0 0 0 1 1/D

0 0 -1 0 0 0 D 1 0 0 0 0
perspective projection perspective-projection

Perspective Space

The homogeneous perspective transformation transforms Euclidean points in 3 (represented as

homogeneous points) to new homogeneous points, which may then be converted to Euclidean

space. These transformed points exist in perspective space. We now examine the properties of the

perspective transformation in terms of the relation between perspective space and the

untransformed object space.

In an affine transformation matrix, the last column is [0 0 0 1]T; multiplication with a direction

vector [a, b, c, 0] yields another direction vector [a, b, c, 0]. In a perspective transform,

however, the last column is [0 0 -1/D 1]T; multiplication with a direction vector yields [a, b, c,

-c/D]. This is a homogeneous point, not direction, and is known as the vanishing point for the

given direction. A set of three-dimensional parallel lines will intersect at the same vanishing point,

with the one exception that parallel lines in the xy plane remain parallel.

In many graphics systems, points in object space are transformed by a matrix that is the

concatenation of all rotation, scaling, translation, and perspective transformations. Usually the

projection transformation is not incorporated into this concatenation, for several reasons. First, the

perspective matrix is invertible whereas the perspective-projection matrix is singular. Secondly,

the value of z before projection is often useful for incremental scanline depth sorting or for

intensity variation in line drawings. More importantly, however, z can simplify three-

dimensional clipping of Euclidean lines and line segments against the viewing frustum by clipping

homogeneously against a parallelepiped. After clipping, points are simply projected to the z = D

image plane and displayed.

 19

The viewing frustum consists of six planes defined in terms of the eye position, the visible portion

of the image plane, and the allowable depths of three-dimensional objects. The viewing frustum

shown in Figure 7 has a visible image sized 2*Sx by 2*Sy, and the near and far clipping planes

arbitrarily set to z = D and z = F.

image
plane top
r
fa (z = F)
Sy right (Sx , Sy , 0) ar (Sx , Sy , 1)
fa
r
Z ne
Sy

8
near
(z = D)
Euclidean space perspective space

Figure 7: viewing frustum

To effect the image size above, we simply scale the first and second columns of the perspective

matrix by Sx and Sy respectively. Assuming a square window, Sx and Sy may be derived from the

field of view, fov, by Sx = Sy = D.tan(fov/2).

To simplify clipping, wed like the near clipping plane to transform to z = 0, and the far clipping

plane to transform to z = 1. By considering the transformations of the points (0, 0, D) and (0, 0,

F), where [P] is the perspective matrix:

[0 0 D 1][P] = [0, 0, D-1, 1] ~ (0, 0, D-1)

[0 0 F 1][P] = [0, 0, F-1, F/D] ~ (0, 0, D-D/F)

we see that the near and far clipping planes can be transformed by scaling the z-terms, resulting in:

[ ]
1/Sx 0 0 0

0 1/Sy 0 0

[P] = 1 .
0 0 1/D
D(1-D/F)

-1
Now, as desired: 0 0 0
1-D/F
 20

[
[0 0 D 1][P] = 0, 0,
D
D(1-D/F)
-
1
1-D/F ]
, 1 = [0, 0, 0, 1]

[
[0 0 F 1][P] = 0, 0,
F
-
1
,
D(1-D/F) 1-D/F D
F
][
= 0, 0,
F-D
F(1-D/F) ]
, 1 = [0, 0, 1, 1].

Example Perspective Transformations

To familiarize ourselves with the perspective matrix, we list the following homogeneous points and

their transformations:

[a b D 1][P] = [a, b, 0, 1] point on near clipping plane (z = D)

transforms to xy plane (z = 0)

[a b F 1][P] = [a, b, 1, 1] far plane transforms to z = 1

[0 0 1 0][P] = [0, 0, 1/(1-D/F), 1] vanishing point for line parallel to z-axis

[a b c 0][P] = [a, b, c, 1] vanishing point for arbitrary line

[a b D 1][P] = [a/Sx, b/Sy, 0, 1] but for scale, point on image plane unchanged

[a b 0 0][P] = [a/Sx, b/Sy, 0, 0] point at infinity in xy plane is unchanged;

i.e, parallel lines in xy plane remain parallel

aD 1
[a 0 b 0][P] = [ , 0, , 1] point at infinity in xz plane lies
bSx 1-D/F

on perspective horizon

[Sx Sy D 1][P] = [1, 1, 0, 1] frustum corners become image window corners

-1
[0 0 0 1][P] = [0, 0, , 0] eye transforms to point at infinity on
1-D/F
 21

negative perspective z-axis

Associated Vertex Transformations

The efficient shading of a polygon whose vertices have been transformed to the display screen is

usually accomplished by incremental techniques described in the literature [Newman]. Similar

techniques may be applied to vertex parameters such as color and texture. Any parameter

interpolated across a polygon must follow the same transformation as the polygon vertices; if the

transformation includes perspective, a homogeneous division is required at each pixel within the

screen space polygon (in practice, linear interpolation in screen space is usually not objectionable

for parameters such as color, but is often apparent when applied to texture coordinates).

Specifically, for perspective transformations, an interpolated parameter will have the form

(asx+bsy+c)/(dsx+esy+f),

where sx and sy are screen space coordinates. [Heckbert] refers to this as rational linear

interpolation, in a detailed discussion of texture coordinate transformations. [Blinn92], which

provides an intuitive review of graphics transformations and homogeneous coordinates, refers to

this as hyperbolic interpolation. In the case where the polygon is parallel to the projection plane,

d = e = 0 and the above equation reduces to linear interpolation.

Homogeneous Clipping

Three-dimensional line segments and polygon edges must be clipped to the viewing frustum. This

should be performed in perspective space, after vertex colors have been computed; otherwise,

clipping, by modifying a vertexs location, would modify its color. By clipping in perspective space

the integrity of the polygon is maintained longer. Also, in homogeneous screen space the

equations for the clipping planes are of the form x = w, whereas they are more complex in object
 22

space. [Blinn78, Blinn91] develop this material in greater detail.

In this section we use * to represent homogeneous coordinates before the conversion to Euclidean

space. That is, [x*, y*, z*, w*] = [x y z 1][P] and (x, y, z) = (x*/w*, y*/w*, z*/w*).

As shown in Figure 7, the viewing frustum is transformed by the perspective matrix to a

parallelepiped in perspective space, and (x, y, z) is clipped such that -1 x, y 1 and 0

z 1. Note that negative z values result in negative w values, with z = z*/w* > 0. Thus,

points behind the eye must be clipped before the homogeneous division. In summary, we test that:

for positive w: for negative w:

-w* x*, y* -- left, bottom -- -w* x*, y*

x*, y* w* -- right, top -- x*, y* w*

0 z* -- near -- 0 z*

z* w* -- far -- z* w*

Let us represent the line segment to be clipped as, p* = p1*+(p2*-p1*), [0, 1], where p1*
and p2* are the perspective-space endpoints [x1*, y1*, z1*, w1*] and [x2*, y2*, z2*, w2*]. The

intersection of the line segment with a clipping plane is then given in terms of :

x1*+w1* x1*-w1*
left = right =
-(w2*-w1*)-(x2*-x1*) (w2*-w1*)-(x2*-x1*)

y1*+w1* y1*-w1*
bottom = top =
-(w2*-w1*)-(y2*-y1*) (w2*-w1*)-(y2*-y1*)

z1* z1*-w1*
near = far =
-(z2*-z1*) (w2*-w1*)-(z2*-z1*)

For example, consider the perspective space line segment from p1* = [2, y1, z1, 2] to p2* = [-1,

y2, z2, 1/2]. For the first point, p1x* > -p1w* (2 > -2); but, for the second point, p2x* < -p2w* (-1 <

-1/2), implying an intersection with the left clipping plane. We compute

 23

left = (2+2)/(-(1/2-2)-(-1-2)) = 8/9.

The intersection with the left clipping plane is now given by

p* = [-2/3, y1*+(8/9)(y2*-y1*), z1*+(8/9)(z2*-z1*), 2/3].

The projected x-coordinate of p* = x*/w* = (-2/3)(3/2) = -1, which is the left boundary of

the viewing frustum.

Conclusion

In this paper we have offered a unified view of homogeneous coordinates within a Computer

Graphics context. First, a brief historical review revealed that, as the understanding of perspective

and projections increased, new coordinate systems were developed to represent the underlying

spaces; one of these systems was the homogeneous coordinate system, which was later seen to

possess properties useful for Computer Graphics.

Next, we formally introduced the homogeneous coordinate system. Its application in two-

dimensional Euclidean space was discussed in some detail; it was shown that affine transformations

can be effected consistently with matrix multiplication, thus simplifying sequences of

transformations, and that the intersection of two-dimensional lines can be performed without

special cases.

Homogeneous coordinates in three dimensions were discussed in greater detail, with particular

attention devoted to perspective transformations. Finally, a method to clip lines with respect to the

viewing frustum was provided.

 24

Acknowledgements

We gratefully acknowledge Jim Blinn, Charlie Gunn, Paul Heckbert, and Ken Shoemake for

numerous suggestions that improved this paper. We also thank the referees for their thorough

reviews and valuable suggestions, many of which we have incorporated into this paper.

References

Aleksandrov, A.D., Kolmogorov, A.N., Lavientv, M.A. Mathematics, its Content, Methods and

Meaning. MIT Press, 1969.

Ahuja, D.V. and Coons, S.A. Geometry for Construction and Display. IBM Systems Journal (3, 4),

1968, pp. 188-205.

Ayres, F. Projective Geometry. Schaum Publishing, New York, 1967.

Barsky, B. Rational Beta-splines for Representing Curves and Surfaces. IEEE Computer Graphics

and Applications 13 (6), 1993, pp. 24-32.

Bez, H.E. Homogeneous Coordinates for Computer Graphics. Computer-Aided Design 15 (6),

Nov., 1983, pp. 361-365.

Blinn, J. A Trip down the Graphics Pipeline: the Homogeneous Perspective Transform. IEEE

Computer Graphics and Applications. May, 1993, pp. 75-80.

Blinn, J. A Homogeneous Formulation for Lines in 3 Space. Proceedings of SIGGRAPH 77. In

Computer Graphics 11 (3), 1977.

Blinn, J. A Trip Down the Graphics Pipeline: Line Clipping. IEEE Computer Graphics and
 25

Applications 11 (1), 1991, pp. 98-105.

Blinn, J. The Algebraic Properties of Homogeneous Second Order Surfaces. SIGGRAPH Course

Notes (Mathematics of Computer Graphics), 1984.

Blinn, J. Hyberbolic Interpolation. IEEE Computer Graphics and Applications 12 (4), 1992, pp.

89-94.

Blinn, J., Newell, M.E. Clipping Using Homogeneous Coordinates. Proceedings of SIGGRAPH 78.

In Computer Graphics 12 (3), 1978.

Brieskorn, E. and Knrrer, H. Plane Algebraic Curves. Birkhuser, Basel, 1986.

Carlbom, I. and Paciorek, J. Geometric Projections and Viewing Transformations. ACM Computing

Surveys 10 (4), 1978.

Coxeter, H.S.M. Projective Geometry (second edition), Springer-Verlag, New York, 1987.

Farin, G. Curves and Surfaces for Computer Aided Geometric Design (second edition), Academic

Press, New York, 1990.

Foley, J.D., van Dam, A., Feiner, S.K., Hughes, J.F. Computer Graphics, Principles and Practice

(second edition). Addison Wesley, Reading, Massachusetts, 1990.

Heckbert, P. Fundamentals of Texture Mapping and Image Warping. Masters thesis, Computer

Science Division, The University of California at Berkeley, May, 1989.

Herman, I. The Use of Projective Geometry in Computer Graphics. Lecture Notes in Computer

Science 564, Springer Verlag, 1991.

Kantani, K. Computational Projective Geometry. Computer Vision, Graphics, and Image

 26

Processing 54, 1991.

Kline, M. Mathematical Thought from Ancient to Modern Times. Oxford University Press, New

York, 1972.

Mbius, F. Gesammelte Werke, vol. 1, pp. 36-49, Die Barycentrische Calcul. Dr. M. Saendig

oHG, Wiesbaden, Germany, 1967.

Newman, W.M. and Sproull, R.F. Principles of Interactive Computer Graphics. McGraw-Hill,

New York, 1979, pp. 491-501.

Oxford University Press. The Compact Edition of the Oxford English Dictionary. Oxford

University Press, New York, 1971.

Patterson, R.R. Projective Transformations of the Parameter of a Bernstein-Bezier Curve. ACM

Transactions on Graphics, 4, 1985, pp. 276-290.

Pellegrini, M. Ray-shooting and Isotopy Classes of Lines in 3-Dimensional Space. Lecture Notes in

Computer Science 511, Springer Verlag, 1991.

Penna, M.A., Patterson, R.R. Projective Geometry and its Application to Computer Graphics.

Prentice Hall, Englewood Cliffs, N.J., 1986.

Plcker, J. Uber ein neues Coordinatensystem. Journal fr die Reine und Angewandte

Mathematick, vol. 5, 1830, pp. 1-36.

Qiulin, D. and Davies, B.J. Surface Geometry for Computer-Aided Design and Manufacture. Ellis

Horwood, Chichester, 1987.

Riesenfeld, R.F. Homogeneous Coordinates and Projective Planes in Computer Graphics. IEEE
 27

Computer Graphics and Applications 1 (1), Jan., 1981, pp. 50-55.

Roberts, L.G. Homogeneous matrix representation and manipulation of n-dimensional constructs.

MS-1405, Lincoln Laboratory, MIT, May, 1965.

Rogers, D.F., Adams, J.A. Mathematical Elements for Computer Graphics (second edition).

McGraw Hill, New York, 1990.

Ryan, P.J. Euclidean and Non-Euclidean Geometry. Cambridge University Press, Cambridge,

1986.

Snyder, J. and Barr, A. Ray Tracing Complex Models Containing Surface Tessellations.

Proceedings of SIGGRAPH 87. In Computer Graphics 21 (3), 1987.

Stolfi, J. Oriented Projective Geometry. Academic Press, 1991.

Veblen, O. and Young, J.W. A Mathematical Science. In The World of Mathematics, J. Newman,

ed. Simon and Schuster, vol. 3, pp. 1696-1707, 1956.