CSci 6974 and ECSE 6966 Math. Tech.

for
Vision, Graphics and Robotics
Lecture 11, February 27, 2006
Projective, Affine and Similarity
Transformations
Overview of Today’s Lecture
• Complete the discussion from Lecture 10 on homogeneous representations,
including the line at infinity and conics.
• Transformations: projective, similarity, affine.
• Undoing distortions

Projective Transformations
The study of transformations and their properties (especially their invariants)
is fundamental to geometry.
• A projective transformation is an invertible mapping that preserves colin-
earity
• Projective transformations are also known as homographies, projectivities
and colineations.
• One may prove (not easily) that projective transformations may be rep-
resented by a 3 × 3 invertible matrix:

x0 = Hx (1)

Note that H is determined up to an arbitrary scale factor so it has 8

degrees of freedom.
• We are interested in these transformations for 2 primary reasons, one
theoretical and one practical:
– Homographies are the most general linear transformation, and rigid
and affine transformations are special cases of homographies.
– As we will prove in a later lectures, the perspective projection of a
planar surface in the world reduces to a plane homography. The com-
position of two homographies is a homography. Thus, two different
images of a planar surface are related by a homography.
– Even though the matrix H is 3×3, this is still a transformation of two-
dimensional objects because everything is done using homogeneous
coordinates.
Example: Computing Projective Transformations
Suppose we have an image of the face of a building that has severe perspective
distortions. If we know, for example, the dimensions of a window or door that
appears in the image, we can compute a transformation that removes these
distortions.
• Suppose the width of the door is w and the height is h. Let the ratio be
r = h/w.
• Find the four corners in the image. Call the locations x1 through x4 .

• Designate four locations for these to map to: (0, 0, 1)T , (1, 0, 1)T , (0, r, 1)T
and (1, r, 1)T . Note that we are creating an arbitrary Euclidean coordinate
in the image.
• These four “correspondences” allow us to compute a unique transforma-
tion, H.
• We apply this transformation to the entire image. It undoes the effects of
the projective distortion.

Projective Transformations of Lines and Conics

• Using the basic result x0 = Hx we can calculate the perspective transfor-
mations of lines and conics.
• Lines transform as l0 = H−T l, where H−T is the inverse of the transpose
of H.
• Conics transformation as C0 = H−T CH−1 .
• Dual conics transformation as C∗ 0 = HC∗ H.
– You are asked to proved this last property as a practice problem.

Transformation Groups
• The set of projective transformations, together with matrix multiplication,
forms a group.
– A group is a set of elements together with a composition operation
satisfying four fundamental properties: (1) associativity, (2) transi-
tivity, (3) existence of an identity element, and (4) existence of a
unique inverse for each element.
• We then naturally explore subgroups, which are simpler versions of pro-
jective transformations.

2
Euclidean and Similarity Transformations: Motivation
• Changes in coordinate systems in the world are generally Euclidean trans-
formations.
• When the image plane of a camera is parallel to a planar surface and
when the camera is moved parallel to or perpendicular to the surface,
the image geometry of objects on the surface is changed by a similarity
transformation.
• Metric properties are retained by similarity transformations.

Euclidean/Similarity Transformation Matrix Structure

 
s cos θ −s sin θ tx
 s sin θ s cos θ ty  (2)
0 0 1
or, more compactly,  
sR t
(3)
0T 1
• There are four degrees of freedom to this transformation.
• Rotation, scaling and translation are combined into a single matrix. This
shows another advantage of using homogeneous coordinates.
• The scale factor s must be positive for this to be a similarity transforma-
tion. If s = 1, it is a Euclidean transformation.
• Angles between vectors are preserved by this transformation.
• Lengths are scaled by s.

Affine Transformations: Motivation

• Affine transformations are intermediate between similarity and projective
transformations (in the group hierarchy).
• The motion of a planar surface can often be modeled to good accuracy
using an affine transformation.

Affine Transformation Matrix Structure

 
a11 a12 tx
a21 a22 ty  (4)
0 0 1
or, more compactly,  
A t
(5)
0T 1

3
 • A must be full rank.
• t is the translation.

Affine Transformation Properties

• There are six degrees to the affine transformation; the full projective trans-
formation has eight.
• Angles and lengths are not preserved.
• Parallel lines remain parallel.
• The ratio of lengths of parallel segments remains fixed.
• The ratio of two areas remains fixed.

Affine Transformation Decomposition

The 2 × 2 affine matrix A may be factored as
 
s 0
A = R(θ)R(−φ) x R(φ). (6)
0 sy

• The notation R(θ) means a 2 × 2 rotation matrix:

 
cos θ − sin θ
R(θ) = . (7)
sin θ cos θ

• The decomposition is easily seen using the SVD (singular value decompo-
sition).
• Both sx and sy are non-zero.

The Line at Infinity

• Recall: The set of all points at infinity in P 2 forms a line whose parameters
are  
0
l∞ = 0 . (8)
1
This is the line at infinity. The associated points at infinity that are on
this line are x∞ = (x1 , x2 , 0)> .
• An affine transformations moves points at infinity, but leaves them at
infinity. In other words, the line at infinity is fixed under affine transfor-
mations.
• Algebraically, this is because the last row of the affine transformation
matrix is (0, 0, 1).

4
 • A transformation is no longer affine when it moves the line at infinity.
• We will look at an example of this in class.

Decomposition of a Projective Transformation Matrix

• A projective transformation matrix H may be uniquely decomposed as
     
sR t K 0 I 0 A vt
H = HS HA HP = = (9)
0T 1 0T 1 vT v vT v
where
– s > 0,
– v has two components, v is a scalar.
– R is a rotation matrix,
– K is upper triangular with det(K) = 1, and
– A = sRK + tvT .
• Observe how the similarity, affine and projective effects are cleanly sepa-
rated.
• The vector v determines the position of the line at infinity.
• Starting from H, the individual matrices and vectors of the factorization
are easily computed using a variety of decomposition (factorization) tech-
nqies we’ve already learned.

Removing Projective Distortions of Planar Surfaces

• We can’t relate measurements in images to physical reality unless we can
undo projective distortion.
• Understanding this leads to deeper understanding of projective transfor-
mations.
– Techniques for building synthetically generated viewpoints, for in-
serting ads in sporting events, and for marking “first-down” lines in
televised American football are based on such an understanding.
• We’ll start by moving from projective to affine and then briefly discuss
moving from affine to similarity, which is a lot harder.
• Much of this is based on the decomposition in equation 9:
     
sR t K 0 I 0 A vt
H = HS HA HP = = (10)
0T 1 0T 1 nvT v vT v
– Undoing projective effects requires finding the line at infinity and
moving it back to infinity.
– Undoing affine effects requires finding K, which equates to measuring
angles in an affine image.

5
Projective to Affine
• Recall the following important facts:
– On a planar surface in the world (in other words prior to projective
distortions) parallel lines intersect on the line at infinity.
– Projections of parallel lines meet at vanishing points.
– Projective transformations preserve colinearity.
• Therefore, we need to

– Identify the image location of two vanishing points. Denote these

points as y1 and y2 .
– Construct the line joining them. This is the image projection of the
line at infinity.
– Find a simple projective transformation that moves this line back to
infinity.
• The result is that we’ve recovered the structure up to an affine transfor-
mation: parallel lines are now parallel.
• Undoing affine effects is a bit harder.

Affine to Similarity: Solution Overview

We have already seen a partial solution with the computation of a projective
transformation that maps imaged points on the corners of a rectangle to a
rectangle. The following discussion generalizes this when there are only lines
and not corners.
• The “circular points” are fixed points of a rigid transformation. Of course
we can not “see them”, but we can measure their effects.
• To do so, we need to know something about conics formed from lines and
dual conics formed from points.

• We form the dual absolute conic from the circular points and note that
the affine transformation of this conic has a particularly simple form, just
involving the matrix K from (9).
• We then form constraints on this dual absolute conic from the image
angles between lines that are supposed to be perpendicular in the world.
It turns out that knowing two sets of perpendicular lines is enough to fully
constrain the matrix K.
We will discuss some of the details, especially the circular points and line conics.
Most of the rest will be left uncovered and you will not be responsible for it.

6
Circular Points
• Consider the similarity matrix
 
sR t
HS = (11)
0T 1

• The eigenvalues of this matrix are 1, eiθ , e−iθ .

• The eigenvectors corresponding to the last two eigenvalues are
   
1 1
i and −i (12)
0 0
These are the circular points.
• All circles include these points, as we can easily see.
• All similarity transformations preserve the circular points.
• This suggests that if we can identify the locations of the circular points in
the image coordinate system, then we can find a transformation moving
them back to their proper location. This would remove affine effects.
• Unfortunately, we can’t measure their locations in any way because they
are complex. Instead, we measure them indirectly through their dual
conic.

Preliminary: Conic Formed from Two Lines, Dual Conics

Formed from Two Points
• Consider two lines in P 2 , with coordinate vectors l and m.
• We can express a quadratic form that is satisfied for only points x on one
or both of these lines:
Clm = lmT + mlT . (13)

• It is straightforward to show that xT Clm x = 0 if and only if x is on line

l or m.
• Clm is a conic, but of rank 2, a seemingly strange best, but an important
one.
• As an example, we will form a conic from the x axis and the line x = y.
• Analogously, from two points x and y we can form a line (dual) conic:
C∗xy = xyT + yxT . (14)
We can show that a line is on this dual conic if it contains one or both of
these points.

7
C∗∞ — The Conic Dual to the Circular Points
• We can form the conic dual to the circular points. The result is
 
1 0 0
C∗∞ = 0 1 0 (15)
0 0 0

Note: we are doing this in a coordinate system that has not been projec-
tively or affinely distorted
• Let’s look at projective transformations of this conic. Start by reversing
the order of the decomposition in equation 9:
  0
t0
   
I 0 K 0 sR t A
H = HP HA HS = = T (16)
vT v 0T 1 0T 1 v0 v0
T
where A0 = sKR, t0 = Kt, v0 = svT KR, and v 0 = vT Kt + v.
– This decomposition, like the one above, is unique. The terms are
different (and a bit more complicated).
• Using this, the effect of a projective transformation on C∗∞ is

C∗∞ 0 = HC∗∞ HT
= HP HA HS C∗∞ HTS HTA HTP
= HP HA C∗∞ HTA HTP
KKT KKT v
 
= . (17)
vT KKT vT KKT v

Note that K is 2 × 2, and v is 2 × 1.

• This matrix has only 4 degrees of freedom because it symmetric, unique
only up to scale, and rank 2.
• If there are no projective distortions, then the matrix reduces to

KKT
 
∗ 0 0
C∞ = . (18)
0T 0

which has only two degrees of freedom.

Angles and C∗∞ 0

T
T
• Let l = l1 l2 l3 and m = m1 m2 m3 be two lines, and suppose
C∗∞ 0 has not been distorted by either an affine or projective transforma-
tion.

8
 • We can measure the angle between these two lines with respect to the dual
absolute conic:
lT C∗∞ m l1 m1 + l2 m2
cos θlm = = 2
(lT C∗∞ l)1/2 (mT C∗∞ m)1/2 (l1 + l22 )1/2 + (m21 + m22 )1/2
(19)
• If we apply a projective transformation to the lines, we can see that this
relation still holds:
T
l0 C∗∞ 0 m0
cos θlm = . (20)
(l0 T C∗∞ l0 )1/2 (m0 T C∗∞ m0 )1/2

• Here, C∗∞ 0 is the parameter matrix of the dual absolute conic after the
transformation, and l0 and m0 are the tranformed vectors of line parame-
ters. The angle cos θlm is the angle in the original, undistorted coordinate
system.

Undoing Affine Distortions — Metric Rectification

• We still can’t “see” the dual absolute conic or the circular points in the
image. We can generate constraints on them, however, and therefore infer
their parameters!
• Here’s one way. Suppose l1 and m1 and l2 and m2 are two sets of perpen-
dicular lines in the undistorted image. Then, following an affine transfor-
mation (remember, we undid the projective transformation by finding the
line at infinity),
T
l01 C∗∞ 0 m01 = 0 (21)
and
T
l02 C∗∞ 0 m02 = 0, (22)
where
KKT
 
0
C∗∞ 0 = , (23)
0T 0
as in 18.
• These reduce to
m011
 
0 0


l11 l12 S (24)
m012
and
m021
 
0 0


l21 l22 S (25)
m022
where S = KKT

• S is clearly symmetric, is positive definite, and has unit determinant.

9
 • This gives enough information to solve for S.
• We can therefore solve for K, which gives us the affine transformation
matrix.

Summary
• Projective transformations: 3 × 3 invertible matrix.
• Projective transformations model image formation for planar surfaces.

• Hierarchy of transformations: Euclidean, similarity, affine and projective.

• A projective transformation may be uniquely decomposed into similarity,
affine, and purely projective transformations.
• “Stratified reconstruction” is used to undo the effects of projective and
affine transformations and therefore recover the original, similiarity struc-
ture in an image:
– Finding the projection of the line at infinity is used to undo projective
effects.
– The transformation of know angles is used to find the absolute dual
conic and from there the affine component of the transformation.
• Here’s a parting thought on the power of projective transformations: cir-
cles and ellipses are equivalent under affine transformations; circles, el-
lipses and hyperbolas are equivalent under projective transformations!

10
Practice Problems and Potential Test Questions
1. Prove that for projective transformation matrix H, line conics transform
as C∗ 0 = HC∗ HT .
2. Show that similarity transformations satisfy the properties of a group.
3. Find the line conic that is dual to the points (0, 0, 1) and (1, 1, 1). What
is the nullspace of this conic? Give its geometric meaning.
4. Suppose we have computed a transformation matrix H and we want to
apply this transformation to an image, I. Sketch a procedure for doing so.
What problems might you run into along the way? Is there any advantage
to using the inverse mapping H−1 ?
5. Prove that any circle contains the circular points. Do this by first forming
the general conic matrix for a circle.
6. Prove that the decomposition in equation 16 is unique if s > 0 and det K =
1.
7. Find the inverse of a similarity transformation as a function of s, R and
t. Give a geometric interpretation in terms of translation, rotation and
scaling.
8. As described above, the 2 × 2 affine matrix A may be factored as
 
s 0
A = R(θ)R(−φ) x R(φ).
0 sy
What happens to the affine transformation if sx = sy ?

Homework 6 Problems
Solutions to the following problems are due at the start of class on Monday,
March 6, 2006.
1. (15 points)
(a) Given n points xi in P 2 , find necessary and sufficient conditions to
show that these points are collinear.
(b) Given n lines li in P 2 , find necessary and sufficient conditions to
show that these lines intersect at a single point.
(c) Show that these two problems are “duals” of each other in the sense
that if you interchange the roles of lines and points in the proofs, the
proofs are identical.
2. (10 points) How many intersection points are there between a line and a
conic? Find them. To do this, let x and y be two points on the line and
let C be the conic matrix. Write the line parametrically, and solve from
there.

11