Active Basis

1
Active Basis for Modeling, Learning and
Recognizing Deformable Template Ying Nian
Wu UCLA Department of Statistics Joint work with
Zhangzhang Si, Haifeng Gong, and Song Chun
Zhu (this version is outdated, see
http//www.stat.ucla.edu/ywu/AB/ActiveBasisMarkI
I.html for the most updated version)
2
Reproducibility page http//www.stat.ucla.edu/ywu
/ActiveBasis Matlab/C code, Data
3

• Deformable template
• Yuille, Hallinan, Cohen, 1989
• key element in object recognition

• Template must be deformable
• Represent and learn the deformable template

4
Deformable template
Recognize the deformable template
5

• Content
• Representation Active basis model
• Algorithm Shared sketch algorithm
• Architecture Sum-max maps

6
Gabor wavelets Daugman, 1985
Olshausen,
Field, 1996
Localized sine and cosine waves
Model for simple cells in primary visual cortex
7
Gabor wavelets
Operation local Fourier transform, edge detection
Representation wavelet sparse coding, Olshausen,
Field, 1996
raw intensities ? composition of strokes
8
Active basis

• A select set of Gabor wavelet elements
• Each element can perturb its location and
  orientation
• These elements form a deformable template

9
Active basis
An active basis can deform to fit multiple
instances
10
Active basis
An active basis can deform to sketch multiple
instances
11
Math notation
12
Matching pursuit
Mallat, Zhang, 1993
single image
Forward stepwise regression for variable selection
Step 3 explain-away local inhibition
a selected element inhibits highly correlated
ones hard inhibition inhibits those
overlapping in both spatial and frequency
domains

13
Shared matching pursuit
multiple images
Parallel forward regressions
Step 2 local maximization, perturb the active
element to sketch a nearby edge Step 3 local
inhibition
14
Shared matching pursuit
Pool statistics over parallel regressions
Approximation Step 3 hard inhibition,
non-overlapping
maximum contrast
15
Pursuit index and template matching score

• Vote for selecting the next element
• h-function monotone increasing
• discounts strong edges in
  background

Log-likelihood scoring Active correlation linear
scoring
non-probabilistic
16
Shared sketch algorithm
--- approximate shared matching pursuit

• Only slightly more complicated than edge
  detection
• Parallel objective oriented edge detection

17
Shared sketch algorithm

• Decreasing order in log-likelihood ratio
  (template matching score)
• Background edges ignored
• Learning takes 4-5 seconds after convolution

18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
Recognizing learned deformable template
SUM1 Is there an edge here? simple cells
Daugman, 1985 MAX1 Is there an edge nearby?
complex cells local maximum pooling,
Riesenhuber and Poggio, 1999 SUM2 Is there a
composite sketch of edges here?
shape filter for template matching Soft scoring
instead of hard decision
23
Template matching by SUM-MAX
SUM2 map at optimal resolution
Multiple resolutions
24
(No Transcript)
25
Sketch at multiple resolutions
26
Classification
Viola and Jones, 2004 adaboost for face
27
Small training sample
28
EM and K-mean Clustering
29
EM and K-mean Clustering
30
EM and K-mean Clustering
Scored by either log-likelihood or active
correlation
31
Learning from non-aligned training images
32
Learning from non-aligned training images

• Given the bounding box of one training image
• Can be relaxed by multiple starting

Motif finding
33
Learning moving template from video sequence
34
Composing multiple templates
Learn bike template
Split bike template to detect and sketch tandem
bike
35
Is there a tandem bike here?
Is there a wheel nearby?
Is there a wheel here?
Is there an edge nearby?
Is there an edge here?
Soft scoring instead of hard decision
36
Where to split the bike template?
Define and learn parts as highly alignable
sub-templates
37
FRAME --- Filters, Random field, And Maximum
Entropy
Zhu,Wu,Mumford,97,98 Zhu,Liu,Wu,00 Wu,Zhu,Liu,00
38
Textures
Zhu,Liu,Wu,00 Wu,Zhu,Liu,00
39
Generic objects line segments, junctions, etc.
40
Guo, Zhu, Wu, 03,07
41
Scaling connection
Wu, Guo, Zhu, 2008
fine
coarse
Central limit theorem Entropy rate increase
42

• Summary
• Representation Active basis model
• Algorithm Shared sketch algorithm
• Architecture Sum-max maps

43

• Key references
• Olshausen, Field, 1996 wavelet sparse coding
• Riesenhuber and Poggio, 1999 local maximum
  pooling
• 
  cortex-like structure
• Viola and Jones, 2004 adaboost for face
• Acknowlegement
• Chuck Fleming, Alan Yuille, Zhuowen Tu, Leo Zhu
• NSF-DMS 0707055, NSF-IIS 0713652
• Lotus Hill Institute