38
6/26/2006 CGI'06, Hangzhou China 1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University, Burnaby, Canada

6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

  • View
    216

  • Download
    2

Embed Size (px)

Citation preview

Page 1: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 1

Sub-sampling for Efficient Spectral Mesh Processing

Rong Liu, Varun Jain and Hao ZhangGrUVi lab, Simon Fraser University,

Burnaby, Canada

Page 2: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 2

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 3: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 3

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 4: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 4

Spectral Applicationsspectral clustering

[Ng et. al., 02]

spectral mesh compression

[Karni and Gotsman, 00]

watermarking[Ohbuchi et. al., 01]

spectral mesh segmentation [Liu and Zhang, 04]

face recognitionin eigenspace

[Turk, 01]

spectral meshcorrespondence[Jain and Zhang, 06]

“affinity matrix” W,its eigen-decomposition

texture mapping using MDS

[Zigelman et. al., 02]

Page 5: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 5

Spectral Embedding

W = 0.56

j

i

i

j

W = EΛET

n points, dimension 2

1 nE =

1e ne…

embedding space, dimension n

row i

i

j

22/ eWijjiD

Page 6: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 6

Bottlenecks

Computation of W, O(n2) .

Apply sub-sampling to compute partial W.

Eigenvalue decomposition of W, O(n3).

Apply Nyström method to approximate the eigenve

ctors of W.

How to sample to make Nyström work better ?

Page 7: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 7

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 8: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 8

Sub-sampling

Compute partial affinities

n points

O (n2)

O (l . n)complexity:Z = X U Y

l sample points

W =

affinities between X and Yaffinities within

X

Page 9: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 9

Nyström Method [Williams and Seeger, 2001]

Approximate Eigenvectors

W =

A

BT

B

C, A = UΛUT

O (n3)

O (l2 . n)

complexity:

U =

U

BTUΛ-1

approximate eigenvectors

Page 10: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 10

Schur Complement

U

BTUΛ-1

ΛU

BTUΛ-1

T

=

A

BTA-1B

B

BT

W = UΛUT =

W =

A

C

B

BT

Schur Complement = C - BTA-1BF F

Practically, SC is not useful to measure the quality of a sample set.

SC =

Page 11: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 11

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 12: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 12

PCA and KPCA [Schölkopf et al, 1998]

covariance matrix CX

dimension 2

covariance matrix Cφ(X)

X

)(

feature space, high dimension (infinite)

)( X

is implicitly defined by a kernel matrix K, where Kij= ji ,

Page 13: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 13

Training Set for KPCA

)(

K =

L

MT

M

N

L = EΛET

E =

E

MTEΛ-1

˙Λ-1/2

Page 14: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 14

Nyström Method and KPCA

W =

A

BT

B

C

A = UΛUT

U =

U

BTUΛ-1

Nyström

KPCA w/ training

set

K =

L

MT

M

N

L = EΛET

E =

E

MTEΛ-1

˙Λ-1/2

Page 15: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 15

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 16: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 16

When Nyström Works Well ? When the training set of KPCA works well ?

1 2

3

4

5

]||||[ 54321 P22 ||)(||||)(|| j

T

jj yPPy

j

jTraining set should minimize:

subspace spanned by training points

)( jy

Page 17: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 17

Objective Function

)(tr 1BABT

j

jminimize:

maximize:

W =

A

BT

B

C

)( 32 lmlO evaluation:

Page 18: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 18

Compare Γ and SCGiven two sampling sets

S1 and S2

,21 12 SCSC

1. Test data are generated using Gaussian distribution;

2. Test is repeated for 100 times;

3. 4% inconsistency.

Page 19: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 19

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 20: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 20

How to sample: Greedy Scheme

Maximize:

Greedy Sampling Scheme:

)(tr 1BABT W =

A

BT

B

C

A B

Best candidate sampling scheme:To find the best 1% with probability 95%, we only need to search for the best one from a random subset of size 90 (log(0.01)/log(0.95)) regardless of the problem size.

Page 21: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 21

Properties of Γ

)(

)(tr

)(tr

)(tr

1

1

1

1

11

11

A

A

BBA

BAB

T

T

T

T

TTBB 11

(0, m), m is the column size of B

maximize 1T(A-11)1. A is symmetric.

2. Diagonals of A are 1.

3. Off-diagonals of A are in (0, 1).

It can be shown that when A’s columns are

canonical basis of the Euclidean space, the

maxima is obtained.

Page 22: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 22

How to Sample: Farthest Point Scheme

A = 1

1

1

In order for A’s columns to be close to canonical basis, the off-diagonals should be close to zero.

This means the distances between each pair of samples should be as large as possible, namely

22/ jiD

ij eA

Samples are mutually farthest away.

Page 23: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 23

Farthest Sampling Scheme

Page 24: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 24

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 25: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 25

Mesh Correspondence

M(1) D(1) W(1) EΛ-1/2M(1)

M(2) D(2) W(2) EΛ-1/2 M(2)

Page 26: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 26

without sampling

farthest point sampling

random sampling

(vertices sampled: 10,

total vertices: 250)

Page 27: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 27

(vertices sampled: 10

total vertices: 2000)

Page 28: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 28

correspondence error against mesh size

• correspond a series a slimmed mesh with the original mesh

• a correspondence error at a certain vertex is defined as the geodesic distance between the matched point and the ground-truth matching point.

Page 29: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 29

Mesh Segmentation

M D W EΛ-1/2

Page 30: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 30

• (b, d) obtained using farthest point sampling

• (a, c) obtained using random sampling

• faces sampled: 10

• number in brackets: value of Γ

Page 31: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 31

w/o sampling, it takes 30s to handle a mesh with 4000 faces.

2.2 GHz Processor

1GB RAM

Page 32: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 32

Roadmap

Background Nyström Method Kernel PCA (KPCA) Measuring Nyström Quality using KPCA Sampling Schemes Applications Conclusion and Future Work

Page 33: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 33

Conclusion

Nyström approximation can be considered as using training data in Kernel PCA.

Objective function Γ effectively quantifies the quality of a sample set.

Γ leads to two sampling schemes: greedy scheme and farthest point scheme.

Farthest point sampling scheme outperforms random sampling.

Page 34: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 34

Future Work

Study the influence of kernel functions to Nyström method.

Further improve the sampling scheme.

Page 35: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 35

Thank you !

Questions ?

Page 36: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 36

Mesh Correspondence

1. Given any two models, M(1) and M(2), build the geodesic distance matrices D(1) and D(2). Dij encodes the geodesic distance between vertices i and j;

2. D(1) W(1) , D(2) W(2) , using Gaussian kernel.

3. Compute the eigenvalue decomposition of W(1) and W(2), and use the corresponding eigenvectors to define the spectral-embedded models M(1) and M(2).

handle bending, uniform scaling and rigid body transformation.

4. Compute the correspondence between M(1) and M(2).

Page 37: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 37

Mesh Segmentation

1. Given a model M, somehow define the distances between each pair of faces; the distances are stored in matrix D;

2. D W ;3. Compute the eigenvalue decomposition of W, and use the

eigenvectors to spectral-embed the faces.4. Cluster (K-means) the embedded faces. Each cluster corre

sponds to a segment of the original model.

Page 38: 6/26/2006CGI'06, Hangzhou China1 Sub-sampling for Efficient Spectral Mesh Processing Rong Liu, Varun Jain and Hao Zhang GrUVi lab, Simon Fraser University,

6/26/2006 CGI'06, Hangzhou China 38

Maximize:

Given any two sampling sets S1 and S2 , S1 is superior to S2 iff

Efficient to compute.

Minimize: (schur complement)

S1 is superior to S2 iff

Very expensive to compute.

Γ and Schur Complement

)(tr 1BABT

SC = C - BTA-1B

21 SS

21SCSC SS