CVPR2010: Advanced ITinCVPR in a Nutshell: part 5: Shape, Matching and Divergences

Post on 04-Jul-2015

187 views 2 download

Transcript of CVPR2010: Advanced ITinCVPR in a Nutshell: part 5: Shape, Matching and Divergences

Tutorial

Advanced Information Theory in CVPR “in a Nutshell”

CVPRJune 13-18 2010

San Francisco,CAShape Matching with I-divergences

Anand Rangarajan

Shape Matching with I-Divergences

Groupwise Point-set Pattern RegistrationGiven N point-sets, which are denoted by {X p, p ∈ {1, ...,N}}, thetask of multiple point pattern matching or point-set registration is torecover the spatial transformations which yield the best alignment ofall shapes.

2/29

Problem Visualization

3/29

Problem Visualization

3/29

Group-wise Point-set Registration

Principal Technical ChallengesI Solving for nonrigid deformations between point-sets with

unknown correspondence is a difficult problem.

I How do we align all the point-sets in a symmetric manner sothat there is no bias toward any particular point-set?

4/29

From point-sets to density functions

5/29

From point-sets to density functions

5/29

Group-wise Point-set Registration

From point-sets to density functionsI Point sets are represented by probability density functions.I Intuitively, if these point sets are aligned properly, the

corresponding density functions should be similar.

Question: How do we measure the similarity between multipledensity functions?

6/29

Group-wise Point-set Registration

From point-sets to density functionsI Point sets are represented by probability density functions.I Intuitively, if these point sets are aligned properly, the

corresponding density functions should be similar.

Question: How do we measure the similarity between multipledensity functions?

6/29

Divergence Measures

Kullback-Leibler divergence

DKL(p‖q) =

ˆp(x) log

p(x)

q(x)dx

where p(x), q(x) are the probabilitydensity functions.

J divergenceGiven two probability densityfunction p and q, the symmetric KLdivergence is defined as:

J(p, q) =12

(DKL(p‖q) + DKL(q‖p))

7/29

Divergence Measures

Kullback-Leibler divergence

DKL(p‖q) =

ˆp(x) log

p(x)

q(x)dx

where p(x), q(x) are the probabilitydensity functions.

J divergenceGiven two probability densityfunction p and q, the symmetric KLdivergence is defined as:

J(p, q) =12

(DKL(p‖q) + DKL(q‖p))

7/29

Motivating the JS divergence

Modeling two shapes

X

Y

p(X |θ(1)) =

N1∏i=1

1K1

K1∑a=1

p(Xi |θ(1)a ), p(Y |θ(2)) =

N2∏j=1

1K2

K2∑b=1

p(Yj |θ(2)b )

8/29

Motivating the JS divergence

Modeling the overlay of two shapes with identity of origin

X Y

p(X ∪ Y |θ(1), θ(2)) = p(X |θ(1))p(Y |θ(2))

8/29

Motivating the JS divergence

Modeling the overlay of two shapes without identity of origin

Z

p(Z |θ(1), θ(2)) =N1

N1 + N2p(Z |θ(1)) +

N2

N1 + N2p(Z |θ(2))

8/29

Likelihood Ratio

I Which generative model do you prefer? The union of disparateshapes where identity of origin is preserved or one combinedshape where the identity of origin is suppressed.

I Likelihood ratio:

logΛ = logp(Z |θ(1), θ(2))

p(X ∪ Y |θ(1), θ(2))=

N1N1+N2

p(Z |θ(1)) + N2N1+N2

p(Z |θ(2))

p(X |θ(1))p(Y |θ(2))

I Z is understood to arise from a convex combination of twomixture models p(Z |θ(1)) and p(Z |θ(2)) where the weights ofeach mixture are proportional to the number of points N1 andN2 in each set.

I Weak law of large numbers leads to Jensen-Shannon divergence.

9/29

JS Divergence for multiple shapes

JS-divergence of shape densities

JSπ(P1,P2, ...,Pn) = H(∑

πiPi)−∑

πiH(Pi) (1)

where π = {π1, π2, ..., πn|πi > 0,∑πi = 1} are the weights of the

probability densities Pi and H(Pi) is the Shannon entropy.

10/29

Atlas estimation

Formulation using JS-divergence

JSβ(P1,P2, ...,PN) + λ

N∑i=1

||Lf i ||2

=H(∑

βiPi )−∑

βiH(Pi ) + λN∑

i=1

||Lf i ||2.

f i is the deformation function corresponding to point set X i ;Pi = p(f i (X i )) is the probability density for deformed point-set.

11/29

Multiple shapes: JS divergence

JS divergence in a hypothesis testing framework:I Construct a likelihood ratio between i.i.d. samples drawn from a

mixture (∑

a πaPa) and i.i.d. samples drawn from aheterogeneous collection of densities (P1,P2, ...,PN).

I The likelihood ratio is then

Λ =

∏Mk=1

∑Na=1 πaPa(xk)∏N

a=1∏Na

ka=1 Pa(xaka

).

I Weak law of large numbers gives us the JS-divergence.

12/29

Group-wise Registration ResultsExperimental results on four 3D hippocampus point sets.

13/29

Shape matching via CDF I-divergences

I Model each point-set by a cumulative distribution function(CDF)

I Quantify the distance among cdfs via an information-theoreticmeasure [typically the cumulative residual entropy (CRE)]

I Minimize the dis-similarity measure over the space ofcoordinate transformation parameters

14/29

Havrda-Charvát CRE

HC-CRE: Let X be a random vector in Rd , we define the HC-CREof X by

EH(X ) = −ˆ

Rd+

(α− 1)−1(Pα(|X | > λ)− P(|X | > λ))dλ

where X = {x1, x2, . . . , xd}, λ = {λ1, λ2, . . . , λd}, and |X | > λmeans |xi | > λi , Rd

+ = {xi ∈ Rd ; xi ≥ 0; i ∈ {1, 2, . . . , d}}.

15/29

CDF-HC Divergence

CDF-HC Divergence : Given N cumulative probability distributionsPk , k ∈ {1, . . . ,N}, the CDF-JS divergence of the set {Pk} isdefined as

HC (P1,P2, . . . ,PN) = EH(∑k

πkPk)−∑k

πkEH(Pk)

where 0 ≤ πk ≤ 1,∑

k πk = 1, and EH is the HC-CRE.

16/29

CDF-HC Divergence

Let P =∑

k πkPk

HC (P1,P2, . . . ,PN)

= −(α− 1)−1(

ˆRd

+

Pα(X >λ)dλ−∑k

πk

ˆRd

+

Pαk (Xk>λ)dλ)

=∑k

πk

ˆRd

+

P2k (Xk>λ)dλ−

ˆRd

+

P2(X >λ)dλ (α = 2)

17/29

Dirac Mixture Model

Pk(Xk > λ) =1

Dk

Dk∑i

H i (x, xi )

where H(x, xi ) is the Heaviside function (equal to 1 if allcomponents of x are greater than xi ).

0

10

20

30

40

50

60

70020

4060

80

0

0.5

1

18/29

CDF-JS, PDF-JS & CDF-HC

0 1 2

0

0.5

1

1.5

2

2.5

3

3.5Before Registraion

0 1 2

0

0.5

1

1.5

2

2.5

3

3.5CDF−JS

0 1 2

0

0.5

1

1.5

2

2.5

3

3.5PDF−JS

0 1 2

0

0.5

1

1.5

2

2.5

3

3.5CDF−HC

0 2 4 6 80

2

4Before Registraion

0 2 4 6 80

2

4CDF−JS

0 2 4 6 80

2

4PDF−JS

0 2 4 6 80

2

4CDF−HC

19/29

2D Point-set Registration for CC

0 0.2 0.4 0.6−0.1

0

0.1

0.2

Point Set 1

0 0.2 0.4 0.6−0.1

0

0.1

0.2

Point Set 2

0 0.2 0.4 0.6

0

0.1

0.2

Point Set 3

0 0.2 0.4 0.6−0.1

0

0.1

0.2

Point Set 4

0 0.2 0.4 0.6

0

0.1

0.2

Point Set 5

0 0.2 0.4 0.6

0

0.1

0.2

Point Set 6

0 0.2 0.4 0.6

0

0.1

0.2

Point Set 7

0 0.2 0.4 0.6−0.1

0

0.1

0.2

Before Registration

0 0.2 0.4 0.6

0

0.1

0.2

After Registration

20/29

With outliers

0.2 0.4 0.6 0.8 1

0

0.1

0.2

Before Registration

0.4 0.6 0.8 1 1.2

0

0.1

0.2

After PDF−JS Registration

0 0.2 0.4 0.60

0.1

0.2

After CDF−HC Registration

21/29

With different α values

0.5 1

00.10.2

Initial Configuration

0 0.2 0.4 0.60

0.10.2

α=1.1

0 0.2 0.4 0.60

0.10.2

α=1.3

0 0.2 0.4 0.60

0.10.2

α=1.5

0 0.2 0.4 0.60

0.10.2

α=2

0 0.2 0.4 0.60

0.10.2

α=1.9

0 0.2 0.4 0.60

0.10.2

α=1.7

0 0.2 0.4 0.60

0.10.2

α=3

0 0.2 0.4 0.60

0.10.2

α=4

0 0.2 0.4 0.60

0.10.2

α=5

22/29

3D Point-set Registration for Duck

02040600 100 200

0

50

100

150

Point Set 1

02040600 100 200

0

50

100

150

Point Set 2

02040600 100 200

0

50

100

150

Point Set 3

02040600 100 200

0

50

100

150

Point Set 4

02040600 100 200

0

50

100

150

Before Registration

0500 100 200

0

50

100

150

After Registration

23/29

3D Registration of Hippocampi

0

100

200

0

50

100

0

10

20

Point Set 10

100

200

0

50

100

0

10

20

Point Set 2

0

100

200

0

50

100

0

10

20

Point Set 40

100

200

0

50

100

051015

Point Sets Before Registration0

100

200

0

50

100

0

10

20

Point sets After Registration

0100

200

0

50

100

0

10

20

Point Set 3

24/29

Group-Wise Registration Assessment

The Kolmogorov-Smirnov (KS) statistic was computed to measurethe difference between the CDFs.

I With ground truth1N

N∑k=1

D(Fg ,Fk)

I Without ground truth

K =1

N2

N∑k,s=1

D(Fk ,Fs)

25/29

KS statistic for comparison

Table: KS statistic

KS-statistic CDF-JS PDF-JS CDF-HCOlympic Logo 0.1103 0.1018 0.0324

Fish with outliers 0.1314 0.1267 0.0722

Table: Average nearest neighbor distance

ANN distance CDF-JS PDF-JS CDF-HCOlympic Logo 0.0367 0.0307 0.0019

Fish with outliers 0.0970 0.0610 0.0446

26/29

KS statistic for comparison (contd.)

Table: Non-rigid group-wise registration assessment without ground truthusing KS statistics

Before Registration After RegistrationCorpus Callosum 0.3226 0.0635

Corpus Callosum with outlier 0.3180 0.0742Olympic Logo 0.1559 0.0308

Fish 0.1102 0.0544Hippocampus 0.2620 0.0770

Duck 0.2287 0.0160

27/29

KS statistic for comparison (contd.)

Table: Non-rigid group-wise registration assessment without ground truthusing average nearest neighbor distance

Before Registration After RegistrationCorpus Callosum 0.0291 0.0029

Corpus Callosum with outlier 0.0288 0.0092Olympic Logo 0.0825 0.0022

Fish 0.1461 0.0601Hippocampus 13.7679 3.1779

Duck 15.4725 0.3280

28/29

Discussion

I I-divergences for shape matching avoid correspondence problemI Symmetric, unbiased registration and atlas estimationI Shape densities modeled as Gaussian mixtures, cumulatives

directly estimatedI JS (pdf and cdf-based) and HC divergences usedI Estimated atlas useful in model-based segmentation

29/29