random walker cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · trust-region graph...

33
Random Walker Cosegmentation Maxwell D. Collins Jia Xu Leo Grady Vikas Singh {mcollins,jiaxu}@cs.wisc.edu [email protected] [email protected] August 20, 2012

Upload: others

Post on 13-Aug-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Random Walker Cosegmentation

Maxwell D. Collins Jia Xu Leo Grady Vikas Singh

{mcollins,jiaxu}@cs.wisc.edu

[email protected]

[email protected]

August 20, 2012

Page 2: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Outline

Introduction to Cosegmentation

↓ ↓

RWCosegScale-free cosegmentation with quasiconvexity

Page 3: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Image Segmentation

Parsing the image into its constituent components – i.e.foreground and background

Solution consists of a labeling on pixels.

Page 4: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

MRF Segmentation

Probabilistic model: Markov RandomField over 0/1 labels xp.

minx

∑p

wpxp +∑p∼q

wpqδ(xp, xq)

Solved via graph-cut.

Boykov & Jolly 2001

Page 5: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Normalized Cuts

Balanced graph clustering.

links(FG,BG)

links(FG,V)+

links(FG,BG)

links(BG,V)

Page 6: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Random Walker

Graph-based Segmentation

Similarity graph between adjacentpixels.

“Random Walker” EnergyMinimize Dirichlet energy of labels over graph

Eimage(x) = xTLx =∑i∼j

wij(xi − xj)2

Page 7: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Cosegmentation

DefinitionCosegmentation: Identifying common foreground in two ormore images.

Page 8: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Why use cosegmentation?

Segmentation should be easier with more information.

Compared with independent segmentation, will be moreaccurate and require less user intervention.

But...Models for high-level image structure more difficult to optimize.

Page 9: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Why use cosegmentation?

Segmentation should be easier with more information.

Compared with independent segmentation, will be moreaccurate and require less user intervention.

But...Models for high-level image structure more difficult to optimize.

Page 10: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Why use cosegmentation?

Segmentation should be easier with more information.

Compared with independent segmentation, will be moreaccurate and require less user intervention.

But...Models for high-level image structure more difficult to optimize.

Page 11: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Optimization

Cast segmentation as an optimization over segment labels xi

on each image:

minx,h,h̄

∑i∈images

Eimage(xi) + Emodel(hi, h̄)

s.t. x ∈ X

hi = model built from xi

Eimage: Image data: intensities, edgesEmodel: Foreground vs common foreground model h̄

Page 12: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Optimization for Cosegmentation

Related Work

Rother et. al. 2006 Submodular-supermodular,Trust-region Graph Cuts

Mukherjee et. al. 2009 Half-integralityHochbaum et. al. 2009 Parametric max-flowVicente et. al. 2009 Dual decomposition

Additional work using different classes of models:Discriminitive clustering (Joulin et al 2010)Scale-invariant histogram models (Mukherjee et al 2011)Anisotropic diffusion (Kim et al 2011)

Page 13: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Histograms

One basic global model is a histogram.1 Extract features from each

pixel/pixelColor/intensityTexture (Gabor, Winn et. al. 2005)Orientation

2 Bin pixels3 Count pixels in each bin in segment

Page 14: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Histograms

One basic global model is a histogram.1 Extract features from each

pixel/pixelColor/intensityTexture (Gabor, Winn et. al. 2005)Orientation

2 Bin pixels

3 Count pixels in each bin in segment

Page 15: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Histograms

One basic global model is a histogram.1 Extract features from each

pixel/pixelColor/intensityTexture (Gabor, Winn et. al. 2005)Orientation

2 Bin pixels

3 Count pixels in each bin in segment

Page 16: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Histograms

One basic global model is a histogram.1 Extract features from each

pixel/pixelColor/intensityTexture (Gabor, Winn et. al. 2005)Orientation

2 Bin pixels3 Count pixels in each bin in segment

Bin

Count

Page 17: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

RWCoseg: Cosegmentation via Random Walker

RWCoseg leads to a convex quadratic problem:

minx,h,h̄

∑i

xTi Lixi + λ‖hi − h̄‖2

2

s.t. x ∈ [0, 1]n

hi = Hixi

Eimage(x) = xTLx

Emodel(h, h̄) = λ‖hi − h̄‖22

Model is foreground histogram

Page 18: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Parallelization

RWCoseg is a BoxQP

minx,h,h̄

∑i

xTi Lixi + λ‖hi − h̄‖2

2

s.t. x ∈ [0, 1]n

hi = Hixi

minx1,x2

x1...

xm

T

L1 + λHT1 H1 −λH1

. . ....

Lm + λHTmHm −λHm

−λHT1 . . . −λHT

m λmI

x1...

xm

T

s.t. li ≤ xi ≤ ui, xi ∈ [0, 1]ni

Page 19: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Box-QP

Calculate gradients with:

(Li + HTi Hi)xi = Lxi + HT

i (Hixi).

Resulting operations are highly parallel, suitable for GPU

Solved in parallel through gradient-projection methods.

Page 20: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Box-QP

Calculate gradients with:

(Li + HTi Hi)xi = Lxi + HT

i (Hixi).

Resulting operations are highly parallel, suitable for GPU

Solved in parallel through gradient-projection methods.

Page 21: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Gradient-Projection Conjugate-Gradient (GPCG)

From More & Toraldo 1991

Alternating GP and CG phases.

Gradient Projection: Projected line search on

α→ f (P[x + α∇f ])

Conjugate Gradient: active set held constant

Page 22: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Bin Counts

RWCoseg solution does not depend directly on size of bins.

Graph-cut and QPBO methods have O(m2) terms for m thecount of the largest bin.

Bin

Count

Page 23: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Interactive Cosegmentation

Guide segmentation by constraining some pixels.

Page 24: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Scale-Free Segmentation

Histograms not robust to changes in scale.

Desire property

Emodel(hi, h̄) = Emodel(shi, h̄) ∀s ∈ R>0

Page 25: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Quasiconvexity

Perform normalization:

Emodel = −〈h, h̄〉‖h‖2

leading to a function which is not convex.

Can relax condition on Emodel(h, h̄) that it need only bequasiconvex in h.

f ((1− λ)x1 + λx2) ≤ max{f (x1), f (x2)} ∀x1, x2, λ ∈ [0, 1]

Page 26: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Quasiconvexity

Perform normalization:

Emodel = −〈h, h̄〉‖h‖2

leading to a function which is not convex.

Can relax condition on Emodel(h, h̄) that it need only bequasiconvex in h.

f ((1− λ)x1 + λx2) ≤ max{f (x1), f (x2)} ∀x1, x2, λ ∈ [0, 1]

Page 27: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Quasiconvexity

Not closed under addition−−−−−−−−−−−−−→

See Boyd & Vandenberghe 2004, Bazaraa, Sherali & Shetty 2003

Page 28: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Optimization

Can optimize minx f (x) + g(x) for quasiconvex f , g by solving

P(α) =

argminx

f (x)

s.t g(x) ≤ α

for α ∈ [minx g(x), g (argminx f (x))]

∇f∇g

Page 29: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Lipschitz Bounds

Have reduced the problem to a 1D function

(f + g) ◦ P(α)

which is one-sided Lipschitz.

x

f (x)

x

f (x)

Page 30: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Lipschitz Bounds

Have reduced the problem to a 1D function

(f + g) ◦ P(α)

which is one-sided Lipschitz.

x

f (x)

x

f (x)

Page 31: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Lipschitz Bounds

Lower-bound the function from finitely many samples.

α

(f + g)(x) ≥ ((f + g) ◦ P)(α∗)− τ ∀x

Page 32: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Photographs

Images Foreground Images Foreground

Page 33: Random Walker Cosegmentationpages.cs.wisc.edu/~jiaxu/pub/rwcoseg-slides.pdf · Trust-region Graph Cuts Mukherjee et. al. 2009 Half-integrality Hochbaum et. al. 2009 Parametric max-flow

Introduction RWCoseg Scale-Free Cosegmentation Experiments

Acknowledgments

Work included consultations with Nagesh Adluru and Petru M.Dinu.

NLM training grant 5T15LM007359NIH R21AG034315NSF RI 1116584