data-driven shape analysis --- joint shape matching...

51
Data-Driven Shape Analysis --- Joint Shape Matching II 1 Qi-xing Huang Stanford University

Upload: others

Post on 14-Aug-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Data-Driven Shape Analysis--- Joint Shape Matching II

1

Qi-xing HuangStanford University

Page 2: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Last lecture

X º 0

(Positive) semidefiniteness

Xij = XTj1Xi1 X =

264 Im...XTn1

375 h Im ¢ ¢ ¢ Xn1i

Page 3: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Last lecture

Xii = Im; 1 · i · nsubject to

minimizeP

(i;j)2EkXinputij ¡Xijk1

Xij1 = 1;XTij1 = 1; 1 · i < j · n

X º 0X ¸ 0

ADMM [Boyd et al.11]

Page 4: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Last lecture

Xii = Im; 1 · i · nsubject to

minimizeP

(i;j)2EkXinputij ¡Xijk1

Xij1 = 1;XTij1 = 1; 1 · i < j · n

X º 0X ¸ 0

ADMM [Boyd et al.11]

Page 5: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Deterministic guarantee• Exact recovery condition:

#incorrect corres. per point< algebraic-connectivity(G)/4

Page 6: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Constrained optimization framework

minimize

Subject to

Constraints on X

Symmetricmatrices

minimize

Subject to Constraints on X

Asymmetricmatrices

Page 7: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Rotation

minimize

subject to

[Wang and Singer’13]

Page 8: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

• Aligning large-shape collections

• Consistent functional maps

• SLAM

Outline

Page 9: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Aligning shapes

Page 10: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Sequential approach

Optimized orientations-- in the xy plane

Un-oriented shapes Optimized scaling andTranslation along (x,y,z)

Page 11: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Consistent orientations

States of each shape

maxxi

P(i;j)2G

Áij(xi; xj)

Áij(k; l) = 1 Áij(k; l) = 0

Iterative coordinate ascent[Leordeanu et al 06]

MRF Formulation :

Page 12: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Shape matching – local phase

P(i;i0)2G

d2(Fi(Si);Fi0(Si0))

FFD of each shape

Intractable

Page 13: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Shape matching --- local phase

Pair-wise matching

Page 14: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Shape matching --- local phase

Optimize the FFD for eachshape independently

Optimize the FFD for eachshape independentlymii0k = (Fi(pii0k) + Fi0(qii0k))=2

Page 15: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Matching quality

Benchmark datasets [Kim et al 13]

Page 16: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

• Aligning large-shape collections

• Consistent functional maps

• SLAM

Outline

Page 17: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Starting from a Regular Map

lion → cat

Page 18: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Attribute Transfer via Pull-Back

cat → lion

Page 19: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

The Operator View of Maps

Functions on cat are transferred to lion using F

F is a linear operator (matrix)

from cat to lion

Page 20: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Functional Correspondences

20

“probe functions”“F2F” correspondences are just as naturalas “P2P” correspondences

From

To

Page 21: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Functional representation

Part-based correspondences

Page 22: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Representation of Image Relationship

Point features Segments Descriptors

Indicator functionsDelta functions SIFT flow[Liu et al. 09]

2 RNN: #pixels

Page 23: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Functional correspondence/map• Descriptors correspondence

• Segmentation correspondence

Page 24: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Reduced representation

• Basis of functional space– First M eigenfunctions of the graph Laplacian

• Reconstruct any function with small error (M=30)

Binary indicator function Reconstructed function Thresholdedreconstructed function

…Example basis functions

Page 25: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Approach I

Page 26: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Estimation of Functional Maps• Recover functional maps by aligning image features:

Page 27: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Estimation of Functional Maps• Regularization term:

– Correspond bases of similar spectra– Enforce sparsity of map

Map with regularization Map without regularization

Page 28: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Problems of Pair-wise Maps• Transport a function along a cycle

Page 29: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Problems of Pair-wise Maps• Transport a function along a cycle

Page 30: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Problems of Pair-wise Maps• Transport a function along a cycle

– typically does not go back

Page 31: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Problems of Pair-wise Maps• Transport a function along a cycle

– typically does not go back

• Need to rectify pair-wise maps

Page 32: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Consistency of the Network• Cycle-consistency

– A function transported along any loop should be identicalto the original function

ijX

jkX

kiX

Page 33: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Consistency of the Network• Cycle-consistency

– A function transported along any loop should be identicalto the original function

ijX

jkX

kiX

if

kf jf

ki jk ij i iX X X f f

Page 34: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Enforcing Cycle-Consistency

– Consistency term:

Page 35: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Joint Estimation of Functional Maps• Overall optimization

minX

(i;j)2Gwij(kXijDi ¡Djk1

+ ¹X

1·s;s0·M

³j¸si ¡ ¸s

0j jXij(s; s0)

´2+ ¸kXijYi ¡ Yjk2F)

s:t: Y TY = Im

Page 36: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Joint Estimation of Functional Maps• Overall optimization

minX

(i;j)2Gwij(kXijDi ¡Djk1

+ ¹X

1·s;s0·M

³j¸si ¡ ¸s

0j jXij(s; s0)

´2+ ¸kXijYi ¡ Yjk2F)

s:t: Y TY = Im

When Y is fixed: solving independent pair-wise functional maps

Page 37: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Joint Estimation of Functional Maps• Overall optimization

minX

(i;j)2Gwij(kXijDi ¡Djk1

+ ¹X

1·s;s0·M

³j¸si ¡ ¸s

0j jXij(s; s0)

´2+ ¸kXijYi ¡ Yjk2F)

s:t: Y TY = Im

When X is fixed: exact eigen-decomposition problem

Page 38: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

• Indicator function transportation– Better correspondence between “cows” with map

consistency

Consistent Functional Maps

Source

Target

Withoutconsistency

Withconsistency

Page 39: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Approach II

Page 40: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

• There exists latent basis on each object

• Definition: For each basis vector for each loop

• Equivalent definition:

Consistency

or

Expanded

Page 41: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Map collection matrix is low-rank

Low-rank factorization

Page 42: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

• Pair-wise probe functions

• Formulation

Matrix recovery

Local optimization to recover Y

Page 43: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Illustraction

Page 44: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Consistent functions

Page 45: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Multi-level

= = =

Super-objects

Page 46: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

• Aligning large-shape collections

• Consistent functional maps

• SLAM

Outline

Page 47: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Geometry reconstruction

RegistrationReconstruction

Scanning

Page 48: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Joint registration and reconstruction

Initial FinalHuang, Adams, Wand.Bayesian Surface Reconstruction via Iterative Scan Alignment to an Optimized Prototype,SGP’07.

Page 49: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

Fragment reconstruction

Ephesus, turkey

Page 50: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

SLAM

22K scans, 180 points per scan

Huang and Anguelov.High Quality Pose Estimation by Aligning Multiple Scans to a Latent Map. ICRA'10.

Page 51: Data-Driven Shape Analysis --- Joint Shape Matching IIgraphics.stanford.edu/courses/cs468-14-spring/slides... · 2014-04-24 · Initial Final Huang, Adams, Wand. Bayesian Surface

SLAM

Computer history museum Exploratorium museum