Segmentation Through Optimization

1
Segmentation Through Optimization

• Pyry Matikainen

2

• He who fights with monsters should look to it
  that he himself does not become a monster.
• -Friedrich Nietzsche, Beyond Good and Evil

3
Retroactively justify decisions
Gradient ascent via parameter tweaking
4
What is wrong with this?

• Difficult to use
• Difficult to extend
• Difficult to study

5

• Z. Tu and S. C. Zhu (2002)to the rescue!

and also Ren and Malik (2003)
6

• Z. Tu and S. C. Zhu. Image Segmentation by
  Data-Driven Markov Chain Monte Carlo. PAMI,
  vol.24, no.5, pp. 657-673, May, 2002

The DDMCMC paradigm combines and generalizes
these all other segmentation methods in a
principled way.
7
(No Transcript)
8
Everything is search.
9
(No Transcript)
10
Optimizer
11
What is a good segment?
Ren and Malik (2003)
12
(No Transcript)
13
How do we model a segment?
14
(No Transcript)
15
(No Transcript)
16
(gaussian)
(histogram)
(gabor)
(Bezier)
17
(No Transcript)
18
Region appearance model complexity
Region area
Region perimeter length (smoothness)
Number of regions
Notably absent the data
19
(No Transcript)
20
Brightness
Superpixels (normalized cuts)
Texture (textons)
Oriented energy
21

G(WI)
Classifier
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
Optimizer
26
(No Transcript)
27
MCMC is a technique for sampling from
distributions.
28
(No Transcript)
29
Number of regions
Region
Region?
?
?
?
30
Merge
Split
Boundary competition
Model adaptation
Switching image models
31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34

• Data driven do some clustering to make the MCMC
  faster.

35
Optimizer
36
Tu Zhu
37
(No Transcript)
38
Ren Malik
39
(No Transcript)
40
(No Transcript)
41
Tu Zhu Ren Malik
New paradigm?
Combines and generalizes other techniques?
Principled?
Good results?
1/2
1/2
1/2
0
0
0
1
1/3
42
Optimizer
43
(gaussian)
(mixture of gaussians)
(3x Bezier spline)
44
(g1)
(gaussian)
(g2)
(histogram)
(g3)
(gabor filter)
(g4)
(Bezier spline)
45
Number of regions
Region appearance model parameters
Region appearance model
Pixels in region
46
MCMC
47
Xiaofeng Ren and Jitendra Malik. Learning a
Classification Model for Segmentation. ICCV 2003.
48
Boundary between i and j
49
Classification certainty

• Tu and Zhu 2002
• Sampling P(WI)
• Generative models
• Pixels

• Ren and Malik 2003
• Maximizing G(WI)
• Discriminative models
• Superpixels