CSE 554
Lecture 9: Laplacian Deformation
Fall 2013
CSE554
Laplacian Deformation
Slide 1
Review
Source
• Alignment
– Registering source to target by
Input
rotation and translation
Target
• Rigid-body transformations
• Methods
– Aligning principle directions (PCA)
After PCA
– Aligning corresponding points (SVD)
– Iterative improvement (ICP)
• Combines PCA and SVD
After ICP
CSE554
Laplacian Deformation
Slide 2
Non-rigid Registration
• Rigid alignment cannot account for shape variance
• Non-rigid deformation can give a better fit
Source
Target
Rigid alignment
CSE554
After non-rigid deformation
Laplacian Deformation
Slide 3
Non-rigid Registration
• A minimization problem
– Minimizing the distance between the
deformed source and the target
• “Fitting term”
– Minimizing the distortion to the source
shape
• “Distortion term”
CSE554
Laplacian Deformation
Slide 4
Intrinsic vs. Extrinsic
• Intrinsic methods
– Deforms points on the source curve/surface
– App: boundary curve or surface matching
• Extrinsic methods
– Deforms all points on and interior to the source curve/surface
– App: image or volume matching
CSE554
Laplacian Deformation
Slide 5
Laplacian-based Deformation
• An intrinsic method
– Simple, efficient, producing reasonable results
• Preserving local shape features
– Widely used in graphics applications for interactive deformation
Reference: “Laplacian surface editing”, by Sorkine et al., 2004 (citation ~ 500)
CSE554
Laplacian Deformation
Slide 6
Setup
• Input
– Source with n points: p1,…,pn
• Let the first m points be “handles”
– Target location of handles: q1,…,qm
When deforming the source
to fit a target shape: m=n and
qi is the point on the target
closest to pi.
• Output
– Deformed locations of source points: p1’,…,pn’
q2
Deformed
Source
p3=q3
p1=q1
p2
An example with 3 target points, two of which are stationary (red)
CSE554
Laplacian Deformation
Slide 7
Overview
• Finding deformed locations pi’ that minimize:
E
Ef
Ed
– Ef: fitting term
• Measures how close are the deformed handles to the target
– Ed: distortion term
• Measures how much the source shape is changed
CSE554
Laplacian Deformation
Slide 8
Fitting Term
• Sum of squared distances to target handle locations
m
Ef
pi '
qi
2
i 1
q2
p2
CSE554
Laplacian Deformation
Slide 9
Distortion Term
• Q: How to measure shape?
• A: By “bumpiness” at each vertex
– Laplacian: vector from the centroid of neighbors to the vertex
• Recall that in fairing, we reduced this vector to “smooth out” bumps
• A linear operator over point locations
L pi
pi
1
Ni
pj
j Ni
where Ni are indices of neighboring
vertices of pi
CSE554
pi
Laplacian Deformation
pi1
pi2
pi1
pi2
2
Slide 10
Distortion Term
• Minimizing changes in Laplacians during deformation
– Over all source points
n
Ed
L pi '
i
2
i 1
i: Laplacian at pi before deformation
pi '
pi
L pi '
i
CSE554
Laplacian Deformation
Slide 11
Putting Together
• Finding deformed locations pi’ that minimize:
E
Ef
Ed
m
n
pi '
i 1
qi
2
L pi '
i
2
i 1
– A quadratic equation in terms of variables (pix’, piy’, piz’)
• qi, i are constants
• L[] is a linear operator
CSE554
Laplacian Deformation
Slide 12
Quadratic Minimization
• A general form of quadratic minimization:
k
ai T x
min
bi
2
i 1
– There are s variables: x=(x1,…,xs)T
– Each a1,…, ak is a length-s column vector (linear coefficients)
– Each b1,…, bk is a scalar (constant coefficients)
– k should be greater than s (so that the problem is over-constrained)
CSE554
Laplacian Deformation
Slide 13
Quadratic Minimization
• To solve:
k
ai T x
min
bi
2
i 1
• Re-write in matrix form:
min A x
B
2
a1 T
where A
is a k by s matrix
ak T
b1
B
is a length-k vector
bk
CSE554
Laplacian Deformation
Slide 14
Quadratic Minimization
• The minimizer is where the partial derivatives are all zero
0
Ax
B
x
2 AT A x
2
2 AT B
AT A x
AT B
– To solve for x in this equation:
• Taking matrix inverse (good for small s, but numerically unstable for large s)
x
AT A
1
AT B
• Using specialized linear system solver (LinearSolve in Mathematica,
TNT/LAPACK in C)
CSE554
Laplacian Deformation
Slide 15
Quadratic Minimization
• Re-writing our minimization in the general form
E
Ef
Ed
m
n
pi '
i 1
qi
2
L pi '
i
2
i 1
– In 2D, there are s=2n variables: x = (p1x’,…, pnx’, p1y’,…, pny’ )T
• In 3D, there are s=3n variables
– We will next re-write each quadratic term in 2D as (aix-bi)2
• Can be extended easily to 3D
CSE554
Laplacian Deformation
Slide 16
Quadratic Minimization
• The ai and bi in the fitting term
m
Ef
m
pi '
i 1
qi
2
m
pix '
qi x
2
i 1
pi y '
qi y
2
i 1
– There are 2m quadratic terms
x
2m
ai T x
bi
p1x ', …, pnx ', p2y ', …, pny '
2
i 1
– In the first set of m terms:
• For i=1,…,m, bi=qix, ai contains all zero, except its (i)th entry is 1.
– In the second set of m terms:
• For i=1,…,m, bi+m=qiy, ai+m contains all zero, except its (i+n)th entry is 1
CSE554
Laplacian Deformation
Slide 17
Quadratic Minimization
• The ai and bi in the fitting term
m
Ef
m
pi '
qi
i 1
2
m
pix '
qi x
2
pi y '
i 1
qi y
2
i 1
– There are 2m quadratic terms
2m
ai T x
bi
2
i 1
– Example with 3 vertices and 2 fitting constraints (n=3; m=2):
a1 T
1 0 0 0 0 0
a2 T
0 1 0 0 0 0
a3 T
0 0 0 1 0 0
a4 T
0 0 0 0 1 0
x
p1x '
p2x '
p3x '
p1y '
b1
b2
b3
q1 x
q2 x
q1 y
p2y '
b4
q2 y
p3y '
CSE554
Laplacian Deformation
Slide 18
Quadratic Minimization
• The ai and bi in the distortion term:
n
Ed
L pi
n
L pi '
i
2
ix
2
i 1
– There are 2n quadratic terms
Ni
L piy '
iy
pj
j Ni
2
i 1
2n
ai T x
i 1
– The first set of n terms:
1
n
L pi x '
i 1
pi
bi
x
2
p1x ', …, pnx ', p2y ', …, pny '
• For i=1,…,n, ai is all zero except the (i)th entry is 1, the (j)th entries are -1/|Ni|
for all jNi, and bi=ix
– The second set of n terms:
• For i=1,…,n, ai+n is all zero except the (i+n)th entry is 1, the (j+n)th entries are
-1/|Ni| for all jNi, and bi+n=iy
CSE554
Laplacian Deformation
Slide 19
Quadratic Minimization
• The ai and bi in the distortion term:
n
n
Ed
L pi '
i
L pi x '
– There are 2n quadratic terms
– Example with 3 vertices (n=3):
a2 T
a3 T
a4 T
CSE554
1
2
1
2
1
2
1
2
1
2
1
1
2
2
0 0 0
a6 T
0 0 0
0 0 0
1
2
1
2
1
2
L piy '
1
2
1
2
1
1
2
1
Ni
x
iy
pj
j Ni
2
i 1
2n
ai T x
bi
2
i 1
0 0 0
1 0 0 0
0 0 0 1
T
a5
ix
i 1
1
pi
n
2
i 1
a1 T
L pi
p1x '
p2x '
p3x '
p1y '
b1
b2
b3
b4
1x
p2y '
b5
2y
p3y '
b6
3y
p3
2x
3x
1y
p1
p2
1
Laplacian Deformation
Slide 20
Summary
• Compute Laplacians (i)
• Construct coefficients (ai, bi)
– Put them into matrices (A,B)
• Solve (x)
– AT A x
CSE554
AT B
x
p1x ', …, pnx ', p2y ', …, pny '
Laplacian Deformation
Slide 21
Results
Deformed
A small deformation
CSE554
Laplacian Deformation
Slide 22
Results
Deformed
A larger deformation
CSE554
Laplacian Deformation
Slide 23
Results
Deformed
Stretching
CSE554
Laplacian Deformation
Slide 24
Results
Deformed
Shrinking
CSE554
Laplacian Deformation
Slide 25
Results
Deformed
Rotation
CSE554
Laplacian Deformation
Slide 26
Discussion
• Limitations
– Local features are “skewed”, and they don’t scale with the model
• Reason: Laplacian changes with rotation or scale
– Two bumps that differ by rotation or scale have different Laplacians
• Which will be penalized by our distortion term
pi '
pi
L pi '
L pi
CSE554
L pi
L pi '
Laplacian Deformation
Slide 27
A Better Distortion Term
• Not penalizing rotation and scaling of local features
– Transforming the original Laplacian vectors before comparing to the
deformed Laplacians
n
Ed
L pi '
Ti
i
2
i 1
• Ti is a matrix that describes how the local shape around pi is deformed
pi
CSE554
Ti
Laplacian Deformation
pi '
Slide 28
Key Questions
• How to represent transformations as matrices?
• How to compute Ti?
• We will focus in the derivations of the 2D case
– 3D results will be briefly presented at the end
CSE554
Laplacian Deformation
Slide 29
Transformation Matrices (2D)
• Homogeneous coordinates
– A 2D point: (x,y,1)
• A 2D vector: (x,y,0)
– A 3D point: (x,y,z,1)
• A 3D vector: (x,y,z,0)
CSE554
Laplacian Deformation
Slide 30
Transformation Matrices (2D)
• Translation
– Cartesian coordinates: vector addition
p'x
p'y
vx
vy
px
py
– Homogeneous coordinates: matrix product
p'x
p'y
1
CSE554
1 0 vx
0 1 vy
0 0 1
px
py
1
Laplacian Deformation
Slide 31
Transformation Matrices (2D)
• Isotropic scaling
– Cartesian coordinates: vector scaling
p'x
p'y
s
px
py
– Homogeneous coordinates: matrix product
p'x
p'y
1
CSE554
s 0 1
0 s 1
0 0 1
px
py
1
Laplacian Deformation
Slide 32
Transformation Matrices (2D)
• Rotation
– Cartesian coordinates: matrix product
p'x
p'y
Cos
Sin
px
py
Sin
Cos
– Homogeneous coordinates: matrix product
p'x
p'y
1
CSE554
Cos
Sin
0
Sin
Cos
0
Laplacian Deformation
0
0
1
px
py
1
Slide 33
Transformation Matrices (2D)
• Summary of elementary similarity transformations
– To combine transformations: take the product of these matrices
Trs v
p'x
p'y
1
M
px
py
1
Scl s
Rot
CSE554
Cos
Sin
0
Laplacian Deformation
1 0 vx
0 1 vy
0 0 1
s 0 1
0 s 1
0 0 1
Sin
Cos
0
0
0
1
Translation
by vector v
Scaling by
scalar s
Rotation
by angle

Slide 34
Similarity Transforms (2D)
• General similarity transformations
T
a w tx
w a ty
0 0 1
– The product of any set of elementary matrices can be written this way
– Any choice of (a, w, tx, ty) can be written as a sequence of rotation,
isotropic scaling and translation
a w tx
w a ty
0 0 1
Trs tx , ty .Scl
a2
w2 .Rot ArcTan
a
w
• Note that a and w can’t be both zero
CSE554
Laplacian Deformation
Slide 35
Computing Ti (2D)
• Suppose we know the deformed locations pi’
• Compute Ti as the similarity transform that best fits the
neighborhood of pi to that of pi’
min
Ti pi
pi '
2
Ti pj
pj '
2
j Ni
pi
CSE554
Ti
Laplacian Deformation
pi '
Slide 36
Computing Ti (2D)
• Suppose we know the deformed locations pi’
• Compute Ti as the similarity transform that best fits the
neighborhood of pi to that of pi’
min
Ti pi
pi '
2
Ti pj
pj '
2
j Ni
• This is a quadratic minimization problem for entries of Ti
– E.g., a, w, tx, ty
CSE554
Laplacian Deformation
Slide 37
Computing Ti (2D)
• The matrix form of the minimization is:
min C
a
w
tx
ty
pix '
piy '
pi1x '
pi1y '
pi x
where C
piy
1 0
pi y
pix
0 1
pi1x
pi1y
1 0
pi1y
CSE554
2
pi1x 0 1
is a 2|Ni|+2 by 4 matrix, and Ni={i1,
i2,…} are indices of neighboring
vertices of pi
Laplacian Deformation
Slide 38
Computing Ti (2D)
• By quadratic minimization:
a
w
tx
ty
pix '
piy '
CT C
1
CT
pi1x '
pi1y '
– Linear expressions of variables (pix’ , piy’)
CSE554
Laplacian Deformation
Slide 39
Distortion Term (2D)
• Two parts of each distortion term:
L pi '
Ti
i
2
– Transformed Laplacian:
Ti
i
a
w
tx
ty
D
pix '
piy '
D CT C
1
CT
pi1x '
pi1y '
where D
ix
iy
iy
ix
0 0
0 0
– Laplacian of the deformed locations:
pix '
piy '
L pi '
CSE554
L
pi1x '
pi1y '
1 0
where L
0 1
1
Ni
0
0
1
Ni
Laplacian Deformation
...
...
is a 2 by
2|Ni|+2 matrix
Slide 40
Distortion Term (2D)
• Putting together:
n
Ed
L pi '
Ti
i
2
i 1
pix '
piy '
n
H 1
i 1
pi1x '
pi1y '
2
pi x '
pi y '
n
H 2
i 1
pi1x '
pi1y '
2
where H
L
D CT C
1
CT
and H 1 , H 2 are its rows
– They form 2n quadratic terms (aix-bi)2 for x = (p1x’,…, pnx’, p1y’,…, pny’ )T
• All bi are zero
• Each ai can be extracted from H
CSE554
Laplacian Deformation
Slide 41
Results (2D)
Old distortion
term
New distortion
term
CSE554
Laplacian Deformation
Slide 42
Results (2D)
Old distortion term
CSE554
Laplacian Deformation
New distortion term
Slide 43
Results (2D)
Old distortion
term
New distortion
term
CSE554
Laplacian Deformation
Slide 44
Results (2D)
Old distortion
term
New distortion
term
CSE554
Laplacian Deformation
Slide 45
Registration
• Use nearest neighbors as corresponding target locations
– Assuming the source is already close to the target
• Iterative closest point (ICP)
– 1. For each point on the source pi, treat it as a handle, and assign its
closest point on the target as its target location qi. Compute Laplacianbased deformation.
• A threshold on the closest distance can be used to throw away unlikely
correspondences
– 2. Repeat step (1) until a termination criteria is met.
• Maximum iteration or minimum RMSD improvement
CSE554
Laplacian Deformation
Slide 46
Result
CSE554
After rigid alignment
1 iteration of Laplacian
7 iterations of Laplacian
Overlaying all curves
Laplacian Deformation
Slide 47
Result
• Weighting the distortion term
E
Ef
w Ed
large w
medium w
small w
CSE554
Laplacian Deformation
Slide 48
Similarity Transforms (3D)
• Elementary transformation matrices
– To perform a sequence of transformations: take the product of these
matrices
Trs v
p'x
p'y
p'z
1
M
px
py
pz
1
Scl s
Rot X,
CSE554
1
0
0 Cos
0 Sin
0
0
Laplacian Deformation
1
0
0
0
s
0
0
0
0
1
0
0
0
s
0
0
0
Sin
Cos
0
0 vx
0 vy
1 vz
0 1
0
0
s
0
1
1
1
1
0
0
0
1
Translation
by vector v
Scaling by
scalar s
Rotation by
angle 
around X
axis
Slide 49
Similarity Transforms (3D)
• General similarity transformations in 3D
T
s
h3
h2
0
h3
s
h1
0
h2 tx
h1 ty
s tz
0
1
– Approximates the product of a set of elementary matrices
• Up to a small rotation angle
• May introduce skewing for large rotations
CSE554
Laplacian Deformation
Slide 50
Computing Ti (3D)
• Assuming known deformation, by quadratic minimization:
s
h1
h2
h3
tx
ty
tz
CT C
1
CT
pix '
piy '
pix
0
piz
piy
piz
0
pix
0 1 0
piz '
pi1x '
pi1y '
piz
piy
pix
0
0 0 1
pi1x
0
pi1z
pi1y
pi1z
0
pi1x
0 1 0
pi1x
0
0 0 1
where C
pi1z '
pi1z
pi1y
piy
1 0 0
pi1y 1 0 0
– Linear expressions of the deformed points pi’
• C is a 3|Ni|+3 by 7 matrix
CSE554
Laplacian Deformation
Slide 51
Distortion Term (3D)
• Constructing transformed Laplacian:
Ti
i
D
pix '
piy '
s
h1
h2
h3
tx
ty
tz
pi1z '
ix
where D
iy
iz
CSE554
1
D CT C
piz '
where
CT pi1x '
pi1y '
0
iz
iy
iz
0
ix
iy
1 0 0
ix
0 1 0
0
0 0 1
Laplacian Deformation
Slide 52