31
1 arized partial similarity of shapes NORDIA – CVPR 200 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael Bronstein © 2008 All rights reserved. Web: tosca.cs.technion.ac.il

1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

  • View
    216

  • Download
    0

Embed Size (px)

Citation preview

Page 1: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

1Regularized partial similarity of shapes NORDIA – CVPR 2008

Not only size matters:Regularized partial similarity of shapes

Alexander Bronstein, Michael Bronstein© 2008 All rights reserved. Web: tosca.cs.technion.ac.il

Page 2: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

2Regularized partial similarity of shapes NORDIA – CVPR 2008

I am a centaur.Am I human? Am I equine?Yes, I’m partially human.

Yes, I’m partially equine.

Identity crisis

Partial similarity

Page 3: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

3Regularized partial similarity of shapes NORDIA – CVPR 2008

Partial similarity

Shapes are partially similar if they

have

significa

nt

similar

parts

Illustration: Herluf Bidstrup

Page 4: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

4Regularized partial similarity of shapes NORDIA – CVPR 2008

Shape

Shape: metric-measure space

Set of points

Metric

Measure

Part: -measurable subset

Represented as indicator function

Part of a shape is a shape

Page 5: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

5Regularized partial similarity of shapes NORDIA – CVPR 2008

Significance

= significance of part

Alternatively:

Many ways to define measure

Area of the part

Normalized area

Statistical weighting a là tf-idf

Page 6: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

6Regularized partial similarity of shapes NORDIA – CVPR 2008

Metric

Many ways to define metric

Restricted metric: as the bird flies

Intrinsic metric: shortest path length

Page 7: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

7Regularized partial similarity of shapes NORDIA – CVPR 2008

Metric

Extrinsic

geometry

Intrinsic

geometry

Invariant to rigid deformation

Invariant to inelastic nonrigid deformation

Page 8: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

8Regularized partial similarity of shapes NORDIA – CVPR 2008

Similarity

Extrinsic similarityHow congruent are and

?

Best rigid alignment

Intrinsic similarityHow similar are and ?

Min. distortion embeddingMinimum-distortion

correspondence

Page 9: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

9Regularized partial similarity of shapes NORDIA – CVPR 2008

Correspondence and (dis)similarity

Correspondence: binary relation

Belongs to some class of correspondences

Rigid correspondences

Non-rigid correspondences

Given correspondence , quantify mismatch

it creates between part of and part of

Dissimilarity between two parts

Page 10: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

10Regularized partial similarity of shapes NORDIA – CVPR 2008

Rigid shape similarity

Closest pointEuclidean isometry

ICP

Page 11: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

11Regularized partial similarity of shapes NORDIA – CVPR 2008

Non-rigid shape similarity

GMDS

Page 12: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

12Regularized partial similarity of shapes NORDIA – CVPR 2008

Multicriterion optimization

Shapes are partially similar if they have significant

similar parts

Simultaneously minimize dissimilarity and maximize

significance

over all possible pairs of parts

Criteria are competing: find a trade-off

Page 13: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

13Regularized partial similarity of shapes NORDIA – CVPR 2008

What is better?

Many small parts… …or one large

part?

Not only size matters!

Page 14: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

14Regularized partial similarity of shapes NORDIA – CVPR 2008

Regularization

Shapes are partially similar if they

have

significant, regular, and similar

partsAdd regularization term

is the degree of regularity of part

Page 15: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

15Regularized partial similarity of shapes NORDIA – CVPR 2008

Boundary length regularization

Irregular Regular

Part regularity inversely proportional to boundary length

Isoperimetric inequality (planar case): for fixed area

(significance),

the most regular shape is a circle.

Page 16: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

16Regularized partial similarity of shapes NORDIA – CVPR 2008

Numerical framework

Minimization over all parts

(binary indicator functions)

is NP-hard!

Relax: replace crisp part

by fuzzy part

Continuous optimization

Page 17: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

17Regularized partial similarity of shapes NORDIA – CVPR 2008

Problem “fuzzification”

Substitute binary with a continuous one

Mumford-Shah regularization on curved surfaces

Similarity: no change

Significance: no change

Regularity: fuzzy part has no boundary!

Page 18: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

18Regularized partial similarity of shapes NORDIA – CVPR 2008

Minimization algorithm

Freeze parts and solve for

correspondence

Freeze correspondence and solve for

parts

1

2Step 1: weighted ICP (rigid) or weighted GMDS (nonrigid)

Step 2: Newton descent

Discretization & numerical details in the paper

Page 19: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

19Regularized partial similarity of shapes NORDIA – CVPR 2008

Rigid partial similarity

SIGNIFICANCE

REG

ULA

RIT

Y

Page 20: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

20Regularized partial similarity of shapes NORDIA – CVPR 2008

Without regularization

With regularization

Page 21: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

21Regularized partial similarity of shapes NORDIA – CVPR 2008

Conclusions

Framework for regularized partial similarity

Control over part significance and regularity

Optimal geometry-driven weighting scheme for ICP and

GMDS

Page 22: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

22Regularized partial similarity of shapes NORDIA – CVPR 2008

Published by Springer

To appear in August 2008

~350 pages

Over 50 illustrations

Color figures

tosca.cs.technion.ac.il

Shameless advertisement

Additional information

Page 23: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

23Regularized partial similarity of shapes NORDIA – CVPR 2008

Backup slides

Page 24: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

24Regularized partial similarity of shapes NORDIA – CVPR 2008

Pareto optimality

INSIGNIFICANCE

DIS

SIM

ILA

RIT

Y

UTOPIA

Attainable set

Pareto frontier

Solution is Pareto optimal if there exist no

with

and

simultaneously.

Page 25: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

25Regularized partial similarity of shapes NORDIA – CVPR 2008

Set-valued dissimilarity

INSIGNIFICANCE

DIS

SIM

ILA

RIT

Y

Pareto frontier = set-valued dissimilarity

Analogy: rate-distortion curve in communication

Page 26: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

26Regularized partial similarity of shapes NORDIA – CVPR 2008

Order relations

BLUE < RED

GREEN ? RED

INSIGNIFICANCE

DIS

SIM

ILA

RIT

Y

Is this crocodile greener than ?

Not total ordering of set-valued dissimilarities

Some dissimilarities are not commensurable

Page 27: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

27Regularized partial similarity of shapes NORDIA – CVPR 2008

Scalar-valued dissimilarity

INSIGNIFICANCE

DIS

SIM

ILA

RIT

Y

Convert Pareto frontier into

scalarFixed significance:

Fixed dissimilarity:

Area under the curve

Distance from utopia:

Page 28: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

28Regularized partial similarity of shapes NORDIA – CVPR 2008

Topology regularization

Euler characteristic as regularity

Gauss-Bonnet identity

No isoperimetric inequality for

general curved surfaces

Same boundary length,

different

number of connected

components

Small boundary length does not

necessarily mean regularity

Page 29: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

29Regularized partial similarity of shapes NORDIA – CVPR 2008

Dave Victoria Horse Centaur Seahorse

Data: tosca.cs.technion.ac.il

Page 30: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

30Regularized partial similarity of shapes NORDIA – CVPR 2008

Nonrigid partial similarity

Partial similarityFull similarity Scalar partial similarity

Human

Human-like

Horse-like

Hu

man

-like

Page 31: 1 Regularized partial similarity of shapes NORDIA – CVPR 2008 Not only size matters: Regularized partial similarity of shapes Alexander Bronstein, Michael

31Regularized partial similarity of shapes NORDIA – CVPR 2008

Regularized partial similarity

INSIGNIFICANCE

DISSIMILARITY

IRREGULARITY