This document discusses various 2D and 3D transformations including translation, rotation, scaling, shearing, and affine transformations. For 2D transformations, it explains how to perform operations like translation, rotation, and scaling using transformation matrices. It also covers recovering affine transformation parameters from point correspondences using least squares optimization. For image warping, it describes using bilinear interpolation to map pixel locations under an affine transformation. The document also introduces basic 3D transformations like translation, scaling, and rotation using 4x4 transformation matrices.
This document discusses various 2D and 3D transformations including translation, rotation, scaling, shearing, and affine transformations. For 2D transformations, it explains how to perform operations like translation, rotation, and scaling using transformation matrices. It also covers recovering affine transformation parameters from point correspondences using least squares optimization. For image warping, it describes using bilinear interpolation to map pixel locations under an affine transformation. The document also introduces basic 3D transformations like translation, scaling, and rotation using 4x4 transformation matrices.
This document discusses various 2D and 3D transformations including translation, rotation, scaling, shearing, and affine transformations. For 2D transformations, it explains how to perform operations like translation, rotation, and scaling using transformation matrices. It also covers recovering affine transformation parameters from point correspondences using least squares optimization. For image warping, it describes using bilinear interpolation to map pixel locations under an affine transformation. The document also introduces basic 3D transformations like translation, scaling, and rotation using 4x4 transformation matrices.
This document discusses various 2D and 3D transformations including translation, rotation, scaling, shearing, and affine transformations. For 2D transformations, it explains how to perform operations like translation, rotation, and scaling using transformation matrices. It also covers recovering affine transformation parameters from point correspondences using least squares optimization. For image warping, it describes using bilinear interpolation to map pixel locations under an affine transformation. The document also introduces basic 3D transformations like translation, scaling, and rotation using 4x4 transformation matrices.
Slides have been modified from Dr.Sohaib and Aamer Zaheers
lectures.
2-D Transformations Scaling Original
Translation
Shear
Affine
Rotation
Projective
Basic 2D Transformations cos sin
1 0 d x 0 1 d y
0 0 1
sin cos 0
0 0 1
1 ex 0 1
0 0
sx 0
0 0 0 1
0 sy 0 1 e y 0
0 0 1 0 0 1 0 0 1
Rotation about an Arbitrary Point
Order of Transformations
Scaled
Rotated
Rotated
Scaled
Original
Original
Hierarchy of 2D Transformations
Length Ratios
Length Ratios Length Cross-Ratios
Recovering Best Affine Transformation Input: we are given some correspondences Output: Compute a1 a6 which relate the images
Recovering Affine Transformation Input: weBest are given some correspondences Output: Compute a1 a6 which relate the images
a1
a2
a3
a4
a5
a6
a7
a8
a9
This is an optimization problem Find the best set of parameters, given the input data
Recovering Affine Transformation Input: weBest are given some correspondences Output: Compute a1 a6 which relate the images
a1
a2
a3
a4
a5
a6
This is an optimization problem Find the best set of parameters, given the input data
Six Unknowns
Least Squares Error Solution
E (a) = ((a x + a y + a x ) + (a x n
' 2 j
' 2 + a y + a
y j j) 5 j 6
j =1
Minimize E w.r.t. a E Compute a , put equal to zero, solve simultaneously i
x 2j j x y j j
xj
j 0
x y x y y y 1 j
2 j
x x y x y y x y 2 j
2 j
x j x 'j 0 j
'
0 yjxj 1
j a 2 0 x 'j a3 = j ' x x y a j j 4 j j j
a ' 5 y y y j j a j j j
6 '
j 1 j y j
Pseudo inverse For an over-constrained linear system Ax = b A has more rows than columns Multiply by AT on both sides ATAx = ATb ATA is a square matrix of as many rows as x We can take its inverse x = (ATA)-1ATb
x1 0
x2
0 x3
0 x 4 0
y1 0 y2 0
1 0 0 x1 1 0 0 x2
0 y1 0 y2
y3 0 y4 0
1 0 0 x3 1 0 0 x4
0 y3 0 y4
0 x1 ' 1 a1 y1 ' 0 a2 x2 '
1 a3 y2 ' = 0 a4 x3 '
1 a5 y3 '
0 a6 x4 '
1 y4 ' x
2D Affine Warping
Warping Inputs: Image X Affine Transformation A = [a1 a2 b1 a3 a4 b2]T
Output: Generate X such that X = AX
Obvious Process: For each pixel in X Apply transformation At that location in X, put the same color as at the original location in X
Problems?
Warping
This will leave holes
Because every pixel does not map to an integer location!
Reverse Transformation
For each integer location in X
Apply inverse mapping Problem?
Will not result in answers at integer locations, in general
Solution: Bilinearly interpolate from 4 neighbors
Interpolation In 1D 3
Use y=mx+c m: slope of the line c: intersect m = 1, c = -2
4 4.3 ?
Substitute x=4.3, => y = 2.3
2D Bilinear Interpolation Four nearest points of (x, y)
(x, y )
(x, y )
(x, y ), (x, y ), (x, y ), (x, y )
( x, y )
where
x = int( x) y = int( y ) x = x +1 y = y +1
(x, y )
(x, y )
Bilinear Interpolation f ' ( x, y ) = x y f ( x, y ) + x y f ( x, y ) + x y f ( x, y ) + x y f ( x, y )
x = x x y = y y x = x x y = y y
(x, y )
(x, y ) x
(x, y )
(x, y )
Warping To apply Translation T of [0,1] Calculate inv(T)
Translation T [0, 1]
(0,0)
(0,1)
(0,0)
(0,1)
(1,0)
(1,1)
(1,0)
(1,1)
I1
I2
For every coordinate (y,x) in I2
Apply inv(T) E.g. (0,0) - 0, -1 :::::: clip it (0,1) - 0, 0 :::::: use value in (0,0)
Warping What if the Translation is [0, 0.7] Translation T [0, 0.7] To apply Translation T of [0,0.7] Calculate inv(T), [0, -0.7]
(0,0)
(0,1)
(0,0)
(0,1)
(1,0)
(1,1)
(1,0)
(1,1)
For every coordinate (y,x) in I2
Apply inv(T) E.g. (0,0) - 0, -0.7 :::::: clip it (0,1) - 0, 0 .3 ::::: Where will you pick value for I2( 0,1)? Use Bilinear Interpolation
I1
I2
Warping Color Images
Same as applying to Gray Scale Images Since Color images consists of More than one 2D matrix In case or RGB image Apply transformation to each layer separately
Bilinear Interpolation Also applied separately
3D Transformations
3D Translation Point in 3D given by (X1 Y1 Z1) Translated by (dx dy dz) X2 = X1 + dx Y2 = Y1 + dy Z2 = Z1 + dz
Translation In matrix form
X 2 1 Y 0 2= Z 2 0
1 0
0 0 dx X 1 1 0 dy Y1 0 1 dz Z1
0 0 1 1 T
Inverse Transformation 1 0 T= 0
0 0 dx 1 0 dy 0 1 dz
0 0 1
T T-1 = I
1 0 T 1 = 0
0 0 dx 1 0 dy 0 1 dz
0 0 1
Scaling Point in 3D given by (X1 Y1 Z1) Scaled by (Sx Sy Sz) X2 = X1 * Sx Y2 = Y1 * Sy Z2 = Z1 * Sz
Scaling In matrix form X 2 Sx 0 0 0 X 1 Y 0 Sy 0 0 Y 2= 1 Z 2 0 0 Sz 0 Z1
1 0 0 0 1 1
Inverse Transformation Sx 0 0 0 0 Sy 0 0
S= 0 0 Sz 0
0 0 0 1
S S-1 = I
1 Sx
0 1 S =
0 0
1 Sy
0 0
1 Sz 0
0 1
Shearing What will these do?
1 0
e 0 0 1 0 0
0 1 0
0 0 1
1 0
0 e 0 1 0 0
0 1 0
0 0 1
1 0
e 0 1 1 0 0 e
0 1 0 0
0 0 1 0
0 0 1 0 0
0 1 0
0 0 1
3D Rotation
Rotation about Z-axis
Z-coordinate will not change Z = Z If we ignore the Z-coordinate, it is 2-D transformation in XY plane
(x2 y2 z2)
X ' cos Y ' sin = Z ' 0
1 0
sin cos 0 0
0 0 X 0 0 Y 1 0 Z
0 1 1
(x1 y1 z1)
Rotation Y
Positive rotation is counterclockwise when looking down the axis of rotation towards origin
(x2 y2 z2)
(x1 y1 z1)
Rotation about Principal Axes
Right hand rule XYZ is a right handed system if: Place your right wrist at the origin curl your fingers from X to Y thumb should point along +ve Z
Rotation about Principal Axes
Y
Rotation about Z
X Z
x Rotation about Y
Rotation about X
y x
z y
X ' = X cos Y sin Y ' = Y cos Z sin Z ' = Z cos X sin
Y ' = X sin + Y cos Z ' = Y sin + Z cos X ' = Z sin + X cos Z'= Z X '= X Y '= Y
Rotation about Principal Axes
X ' = X cos Y sin Y ' = Y cos Z sin Z ' = Z cos X sin Y ' = X sin + Y cos Z ' = Y sin + Z cos X ' = Z sin + X cos Z'= Z X '= X Y '= Y Rotation about X
Rotation about Z cos sin
sin cos 0 0
0 0 1 0
0 0 0
0 1 0 cos
0 sin
0 0
0 sin cos 0
Rotation about Y 0 0 0
cos 0
sin
0 sin 1 0 0 cos 0 0
0 0 0
Concatenation of Rotations
Rotation around X by followed by rotation around Y
by followed by rotation around Z by
R = RZ RY RX
Properties of Rotation Matrix
Rotation Matrices are orthonormal and have a determinant of +1 i.e. R R T = R TR = I
The inverse of a rotation matrix is its transpose
3D rigid motion can be described by a rotation and translation X ' r11 Y ' = r 21 Z ' r31
r12 r22 r32
r13 X Tx r23 Y + Ty r33 Z Tz
or
X ' r11 Y ' r = 21 Z ' r31
1 0
r12 r22 r32 0
r13 TX X r23 Ty Y r33 TZ Z
0 1 1
Properties of Rotation Matrix
r11 R = r21 r31
r12 r22 r32
r13 r23 r33
R has 9 unknowns, but orthonormality provides 6 constraints
0 i k rij rkj = rji rjk =
1 i=k j =1 j =1
Hence the number of degrees of freedom of a 3D rotation are 9
6=3 3
Properties of Rotation Matrices
Any concatenation of rotation matrices also forms a
rotation matrix i.e. the matrix remains orthonormal [Proof?] Quite simple,
A rotation matrix transforms its own rows onto the principal axes r11 r12 r r 21 22 r31 r32
r13 r11 1 r23 r12 = ?0
r33 r13 0
r11 r12 r r 21 22 r31 r32
r13 r31 0 r23 r32 = 0
r33 r33 1
r11 r12 r r 21 22 r31 r32
r13 r21 0 r23 r22 = 1
r33 r23 0
Rotation about Arbitrary Axis
Any complex rotation can be described by a single rotation around an axis n by an angle Therefore, any complex rotation can be described by [n1, n2, n3]T and Still 3 degrees of freedom as n can be taken to be a unit vector without any loss of generality
n12 + n22 + n32 = 1
Rotation about Arbitrary Axis
To rotate about an axis n by an angle
1. Set up rotations such that n rotates onto one of the principal axis [How?] 2. Rotate about that axis by 3. Undo the transformations in step 1
Rotation about Arbitrary Axis
Question: Given an arbitrary 3D rotation matrix, how can we find out the axis n and the angle that represents this rotation? - 0.8256 0.40388 - 0.39404 - 0.20084 - 0.86294 - 0.46367
- 0.5273 - 0.30367 0.79356
Given R on the left, how
can we tell n and ?
Eigenvectors and Values of a Rotation Matrix
3D rotation matrix has eigenvalues of 1, cos + i sin and cos -i sin The eigenvector associated with the real eigenvalue represents the axis of rotation [proof?] We talked about this in the lecture conducted on 17 Nov. 2014 Think about what the Eigen Vectors definition is And what effect Rotation has on the axis of rotation?
Summary: 3D Rotation Matrices
Any 3D rotation matrix can be described as
Rotations about Principal Axes
cos sin
sin cos 0 0
0 0 0 0 1 0
0 1
0 1 0 cos
0 sin
0 0
0 sin cos 0
0 0 0
cos 0
sin
0 sin 1
0 cos 0
About origin, in right handed coordinate system,
counter clockwise when looking towards origin from positive axis
Rotation matrix is orthonormal with
Determinant of +1 and 3 dof Inverse of a rotation matrix is its transpose Concatenation of Rotations is also a rotation IMP: A rotation matrix transforms its own rows onto the principal axes
0 0 0
rotation about an axis n by an angle
To rotate about given axis n by : Rotate axes onto a principal axis
Two ways: by computing principal
rotations or by composing appropriate matrix through cross products Rotate about principal axes and then undo the earlier transformation
To compute n and from a 3D rotation
matrix n is the eigenvector corresponding to the real eigenvalue of 1 can be computed by the other 2 eigenvalues, which are cos i sin
Hierarchy of 3D Transformations
Factorizing Transformations
Opposite of Concatenation of Transformations
Given a transformation matrix, decompose it into a sequence of simpler transformations Example: a1 a 3 0
a2 a4 0
b1 1 0 b1 a1 b2 = 0 1 b2 a3 1 0 0 1 0
a2 a4 0
0 0 1
Question: How to factorize the multiplicative part?
a1 a 3
a2 a4
Is the factorization unique?
Special Case: A is symmetric
Eigen values of a real symmetric matrix are real Its eigenvectors can always be written as orthonormal matrix A =
A= T (Implication?)
A non symmetric real matrix M can be decomposed as
M = U S VT (with U and V being orthonormal, S being a diagonal To compute U S and V, Let A = MMT A = (USVT)(USVT)T A = US2UT VT = (US)-1M
Singular Value Decomposition
Let M be a m-by-n matrix whose entries are real numbers. Then M may be decomposed as M = U S VT where: U is an m-by-m orthonormal matrix S is an m-by-n matrix with non-negative numbers on the main diagonal and zeros elsewhere V is an n-by-n orthonormal matrix
Example http://en.wikipedia.org/wiki/Singular_value_decomposition
Singular Value Decomposition
Implication: We can take the multiplicative part of any transform and describe it as a sequence of a rotation, scaling and another rotation 2D Example: Decomposing an Affine Transformation M = 0.95 0.23 0
0.49 0.89 0
>> U * S * V'
>> [U, S, V] = svd(M(1:2, 1:2))
U= -0.78156 -0.62384
-0.62384 0.78156
S=
ans = 0.95 0.23
0.46 0.02 1
0.49 0.89
1.2904 0 0 0.56789 V= -0.68658 -0.72705
-0.72705 0.68658
Interpretation in terms of angles?
Singular Value Decomposition
Implications: Even a simple shear can be written as a rotationscalingrotation
Try it as an exercise
From Peter Shirley: Fundamentals of Computer Graphics
Summary: 2D and 3D Transformations
Image Registration 2D Transformations
Scaling Shear Rotation Translation
Inverse Transformations Rotation about an arbitrary point Concatenation of transformations Order of transformations Factorization of Transformations
