Sparse Bayesian Learning for Efficient Visual Tracking

1
Sparse Bayesian Learning for Efficient Visual
Tracking

• O. Williams, A. Blake R. Cipolloa
• PAMI, Aug. 2005.

Presented by Yuting Qi Machine Learning Reading
Group Duke University 06/24/2005
2
Overview

• Motivations - an extension of SVT
• Bayesian tracking with RVM
• Overall system
• Experimental results

3
Motivations

• Support vector tracking (SVT) 1
• Training a SVM classifier through the labeled
  image database
• For a given test image, the tracked object region
  is located by maximizing the SVM score.
• Using first-order Taylor expansion, Ifinal is
  the linear transformation of image gradient, Ix
  Iy.

Ifinal correct object region I all possible
regions
u,v motion vector
1 Shai Avidan, Support vector tracking, IEEE
Tran. On PAMI, Aug, 2004
4
Motivations

• Limitations of SVT
• Is the optimization efficient using different
  kernels?
• Is the optimization function suitable?
• Smoothing image gradient may decrease
  performance
• Properties of RVM Tracker
• Fully probabilistic regression for displacement
• Observations of displacement are fused temporally
  with motion prediction
• Online tracking

5
Bayesian Tracking with RVM

• Building a displacement expert-RVM
• Train an RVM to learn the relationship between
  images and motion. For a test image region x, RVM
  returns the displacement
• Mapping from image space to state space.
• 4 dimensional state space
• Translation in x, y, rotation, scaling
• Each dimension building one RVM

6

• Creating training dataset
• Given a seed image I containing the labeled ROI
  ?
• Generating training examples z from I
• Sampling random displacements from a uniform
  distribution
• Corresponding state
• Generating example zi from state u
• Real training examples

7

• Learning the expert
• Given N training examples zi, ti, i1,,N.
• The relationship between subimages zi and
  displacement ti is
• Considering additive processing noise
• Learning
• Posterior is also Gaussian

8

• Tracking with the expert
• Given the test image I, initial state u0
• Get ROI x by sampling I around u0.
• The expert outputs the probability distribution
  of the corresponding displacement
• Assume the state transition probability is
• Plug those into Kalman-Bucy filter for tracking

Gaussian innovation
9

• Tracking algorithm

State predict
Innovation
State update
10
Overall System

• A validator is adopted to achieve the tracking
  robustness.
• Absence of the verification of the tracked object
  triggers a exhaustive search over the input image
  by the classifiers.

11
Face Tracking Results
(1)
(2)
(3)

• Row (1) deformation
• Row (2) occlusion (lost track in the last frame)
• Row (3) lighting variation

12
Hand Tracking Results
Cars Tracking Results
13
Long-term Tracking Results
14
Conclusions

• Develop a tracker using sparse probabilistic
  regression by RVMs.
• RVM can be trained from a single image
  (generating training set).
• Robustness is obtained by the object
  verification.