1 Deformable Template as Active Basis Zhangzhang Si UCLA Department of Statistics Ying Nian Wu, Zhangzhang Si, Chuck Fleming, Song-Ch

 

1  
 
 

Deformable Template as Active Basis 
 

Zhangzhang Si

UCLA Department of Statistics 

Ying Nian Wu, Zhangzhang Si, Chuck Fleming, Song-Chun Zhu  ICCV07 

��the work in this talk is outdated, see

http://www.stat.ucla.edu/~ywu/AB/ActiveBasisMarkII.html

for the updated results in our IJCV paper)

 

CIVS, Statistics Dept. UCLA 

2  
 
 

Design a deformable template to model a set of images of a certain object category. The template can be learned from example images. 

Motivation

 

3  
 
 

Representation:  generative and deformable models

1. Sparse coding [Olshausen-Field 96]
2. Deformable templates [Yuille-Hallinan-Cohen 89]
3. Active contours  [Kass-Witkin-Terzopoulos 87]
4. Active appearance [Cootes-Edwards-Taylor 95]
5. Texton model [Zhu et.al. 02]

 
 

Computation:  learning and pursuit algorithm

1.    Matching pursuit [Mallat and Zhang 93]

2.    HMAX [Riesenhuber-Poggio 99, Mutch-Lowe 06]

3.     Adaboost  [Freund-Shapire 96, Viola-Jones 99] 

Related work

 

 
 

selected from a dictionary of Gabor wavelet elements 

Linear additive image model 

Image reconstruction by matching pursuit. 

4  

Two extensions:

1. Encoding a single image               Simultaneously encoding a set of images;
2. Allow each Gabor wavelet element B_i to locally perturb.

 

location 

scale 

orientation

 

5  
 
 

The active basis model 

��Active��: Local perturbation 

When encoding image I_m, we use

the perturbed version of B_i: 

(Gabor elements represented by bar)

CIVS, Statistics Dept. UCLA 

6  
 
 

Deformable template using active basis 

A car template 

An incoming car image: 

(Gabor elements represented by bar)

 

7  
 
 

Deformable template using active basis 

A car template 

Deformed to fit many car instances

 

8  
 
 

Learning the template: pursuing the active basis 

Example images 

B1 

B3 

B2 

# Gabor elements selected 

q(I): background distribution

(all natural images) 

p(I): pursued model to approximate

   the true distribution.

CIVS, Statistics Dept. UCLA 

9  
 
 

Pursuing the active basis 

MLE: 

(Projected on {B₁,��,B_n}) 

(orthogonality of {B₁,��,B_n})

CIVS, Statistics Dept. UCLA 

10  
 
 

Pursuing the active basis

CIVS, Statistics Dept. UCLA 

11  
 
 

Shared pursuit algorithm

 

12  
 
 

Learning the template: pursuing the active basis 

Car instances 

A car template consisting of  60 Gabor elements

 

 
 

• 37 training images, listed in the descending order of log-likelihood ratio
• 4.3 seconds (Core 2 Duo 2.4GHz) , after convolution

 

Experiment 1: learning an active basis model of vehicle 

13  

template

 

14  
 
 

Experiment 2: learning without alignment 

Active basis pursuit + EM 

Given bounding box for the first example for initialization. 

Iterate:

    - Estimate the bounding boxes using current model.

    - Re-learn the model from estimated bounding boxes.

 

15  
 
 

Experiment 3: learning and clustering  

Learning active basis 

EM clustering

 

 
 

Experiment 4: car detection with active basis model 

map of LLR at optimal scale 

Maximum LLR over scale 

• Scan bounding box over the image at multi-resolutions
• Compute log-likelihood ratio by combining responses from active basis

 

16  

LLR: log likelihood ratio 

LLR: log likelihood ratio

 

17  
 
 

Experiment 5: head-and-shoulder recognition 

Some positives 

Some negatives 

Negatives include various in-door and out door scenes, with and without human 

Human head and shoulders, roughly aligned 

Features: using the same set of Gabor filters. 

43 training positives, 157 training negatives

88 testing positives,   474 testing negatives

 

18  
 
 

Experiment 5: head-and-shoulder recognition

    comparing with Adaboost 

ROC of sigmoid model is a further improvement of the result presented in the paper.

 

19  
 
 

1.  An active basis model as deformable template.

2.  An active bases pursuit algorithm for fast learning. 

Main contributions 

http://www.stat.ucla.edu/~ywu/ActiveBasis.html 

Download

1)  Training and testing images

2)  Matlab and mex-C source codes that reproduce all the experiments in the paper and powepoint.