interactive segmentation with super-labels andrew delong western yuri boykovolga vekslerlena...

38
Interactive Segmentation with Super-Labels Andrew Delong Western Yuri Boykov Olga Veksler Lena Gorelick Frank Schmidt

Upload: joseph-lade

Post on 15-Jan-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

Interactive Segmentation with Super-Labels

Andrew Delong

Western

Yuri BoykovOlga VekslerLena Gorelick Frank Schmidt

Page 2: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

2

Natural Images: GMM or MRF?

MRFare pixels in this image i.i.d.? NO!

Page 3: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

3

Natural Images: GMM or MRF?

GMM

Page 4: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

4

Natural Images: GMM or MRF?

MRF

Page 5: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

5

Natural Images: GMM or MRF?

MRF?

Page 6: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

6

Boykov-Jolly / Grab-Cut

[Boykov & Jolly, ICCV 2001] [Rother, Kolmogorov, Blake, SIGGRAPH 2004]

Page 7: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

7

Boykov-Jolly / Grab-Cut

[Boykov & Jolly, ICCV 2001] [Rother, Kolmogorov, Blake, SIGGRAPH 2004]

Page 8: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

8

Boykov-Jolly / Grab-Cut

GMM

GMM

[Boykov & Jolly, ICCV 2001] [Rother, Kolmogorov, Blake, SIGGRAPH 2004]

Page 9: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

9

• Objects within image can be as complex as image itself

• Where do we draw the line?

A Spectrum of Complexity

MRF?GMM?Gaussian? object recognition??

Page 10: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

10

Single Model Per Class Label

GMM

GMM

Page 11: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

11

Multiple Models Per Class Label

GMM

GMMGMM

GMMGMM

GMM

Page 12: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

12

Multiple Models Per Class Label

MRFMRF

Page 13: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

13

Our Energy ¼ Supervised Zhu & Yuille!

• Zhu & Yuille. PAMI’96; Tu & Zhu. PAMI’02• Unsupervised clustering of pixelsboundary

lengthMDL

regularizer+color

similarity+

Page 14: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

14

Our Energy ¼ Supervised Zhu & Yuille!

• Zhu & Yuille. PAMI’96; Tu & Zhu. PAMI’02boundary

lengthMDL

regularizer+color

similarity+

Page 15: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

15

Interactive Segmentation Example

Page 16: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

16

Boykov-Jolly / Grab Cutsegmentation colour models

Page 17: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

17

Ours

segmentation colour models“sub-labeling”

Page 18: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

18

Main Idea

• Standard MRF:

• Two-level MRF:

object MRF

GMMs GMMs

background MRF

image-level MRF

object GMM background GMM

image-level MRF

unknown number of labels in each group!

Page 19: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

19

The “Super-Pixel” View

• Complex object ¼ group of super-pixels• Interactive segmentation ¼

a“user-constrained super-pixel grouping”

Page 20: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

20

The “Super-Pixel” View

• Why not just pre-compute super-pixels?– boundaries may contradict user constraints– user is helpful for making fine distinctions

• Combine automatic (unsupervised) and interactive (supervised) into single energy

Spatially coherent clustering+ MDL/complexity penalty

+ user constraints= 2-level MRF

Like Zabih & Kolmogorov, CVPR 2004

Label Costs, CVPR 2010

Like Boykov & Jolly, ICCV 2001

Page 21: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

21

Process Overview

user constraintspropose models from current super-labeling1 solve 2-level MRF

via α-expansion2

refine all sub-models3

converged

E=503005E=452288

Boykov-Jolly + unsupervised clustering (random sampling) + iterated multi-label graph cuts (like grab-cut)

Page 22: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

22

Our Problem Statement

• Input: set S of super-labels (e.g. ffg,bgg) constraints g : P ! S [ fanyg

fg

bg

any

Page 23: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

23

Our Problem Statement

• Output: set L of sub-labels sub-labeling f : P ! L model params µ` for each `2L label grouping ¼ : L ! S

¼ ±ff`2

`1`3

GMM `1

white

GMM `2

dark gree

n

light green

Page 24: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

24

Our Energy Functional

• Minimize single energy w.r.t. L, µ, f, ¼

E (L ;µ;¼;f ) =X

p2P

Dp(f p) +X

pq2N

wpqV(f p; f q) +X

`2L

h`±̀(f )

data costs smooth costs label costs

`4

`3 `1

`2

Dp( )̀ =½

¡ lnPr(I pjµ̀ ) if gp = any _ gp = ¼( )̀1 otherwise

forc

es tr

ansiti

on

Page 25: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

25

Our Energy Functional

• Minimize single energy w.r.t. L, µ, f, ¼

E (L ;µ;¼;f ) =X

p2P

Dp(f p) +X

pq2N

wpqV(f p; f q) +X

`2L

h`±̀(f )

data costs smooth costs label costs

pay c2 `between group’

pay c1 `within group’V(¢;¢) 2 f0;c1;c2g

Page 26: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

26

Our Energy Functional

• Minimize single energy w.r.t. L, µ, f, ¼

• Penalize number of GMMs used– prefer fewer, simpler models– MDL / information criterion

regularize “unsupervised” aspect

E (L ;µ;¼;f ) =X

p2P

Dp(f p) +X

pq2N

wpqV(f p; f q) +X

`2L

h`±̀(f )

data costs smooth costs label costs

GMMs GMMs

Page 27: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

27

More Examples

Boykov-Jolly 2-level MRF

Page 28: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

28

More Examples

Boykov-Jolly 2-level MRF

Page 29: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

29

More Examples

Boykov-Jolly

2-level MRF

Page 30: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

30

More Examples

Boykov-Jolly

grad studentsbaby panda

2-level MRF

GMM density for blue model

Page 31: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

31

Interactive Co-segmentation

image collection 2-level MRFBoykov-Jolly

(like “iCoseg”, Batra et al., CVPR 2010)

Page 32: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

32

More ExamplesBoykov-Jolly

2-level MRF

Page 33: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

33

More ExamplesBoykov-Jolly

2-level MRF

Page 34: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

34

Beyond GMMs

GMMs plane

GMMs only GMMs + planes

Page 35: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

35

Synthetic Example

GMM

Boykov-Jolly(1 GMM each label)

GMM

GMMGMM

GMM

2-level MRF (GMMs only)

plane

plane

GMM

2-level MRF (GMM + planes)

• object = two planes in (x,y,grey) space• noise = one bi-modal GMM (black;white)

Page 36: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

36

Synthetic Example

plane

plane

GMM

bla

ckw

hit

e

x

2 planes detected

1 GMM

detected

y

black

white

Page 37: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

37

As Semi-Supervised Learning

• Interactive segmentation ¼ a semi-supervised learning– Duchenne , Audibert, Keriven, Ponce, Segonne.

Segmentation by Transduction. CVPR 2008.

– s-t min cut [Blum & Chawla, ICML’01]– random walker [Szummer & Jaakkola, NIPS’01]

Page 38: Interactive Segmentation with Super-Labels Andrew Delong Western Yuri BoykovOlga VekslerLena GorelickFrank Schmidt TexPoint fonts used in EMF. Read the

38

Conclusions

• GMM not good enough for image ) GMM not good enough for complex objects

• Energy-based on 2-level MRF– data costs + smooth costs + label costs

• Algorithm: iterative random sampling, re-fitting, and ®-expansion.

• Semi-supervised learning of complex subspaces with ®-expansion