Online Multiple Classifier Boosting for Object Tracking Tae-Kyun Kim 1 Thomas Woodley 1 Björn Stenger 2 Roberto Cipolla 1 1 Dept. of Engineering, University.

Online Multiple Classifier Boosting for Object Tracking

Tae-Kyun Kim1 Thomas Woodley1 Björn Stenger2 Roberto Cipolla1

1Dept. of Engineering, University of Cambridge2Computer Vision Group, Toshiba Research Europe

The Task: Object TrackingExample sequence 1

Target appearance changes due to changes in- pose - illumination- object deformation

Example sequence 2

Learning Multi-Modal Representations

- Multi-view face detection [Rowley et al. 98, Schneiderman et al. 00, Jones Viola 03]- Multi-category detection, Sharing features [Torralba et al. 04]

Positive examples

Negative examples

Joint Clustering and Training

K-means clustering

Face cluster 1

Face cluster 2

Positive examples Negative examplesFeature pool

[Kim and Cipolla 08, Babenko et al. 08]

Given:

Set of n training samples with labels number of strong classifiers

Learn strong classifiers:

Combine classifier output with“Noisy OR” function

Map to probabilitieswith sigmoid function

MCBoost: Multiple Strong Classifier Boosting[Kim and Cipolla 08, Babenko et al. 08]

• For given weights, find K weak-learners at t-th round of boosting to maximize

• Weak-learner weights found by a line search to maximize

where

• Sample weight update by AnyBoost method [Mason et al. 00]

MCBoost (continued)

MCBoost: Toy Example 1

Input data MCBoost result (K=3)

Toy Example 2

Standard AdaBoost

MCBoost [Kim and Cipolla 08]

MC Boost with weighting function QMC Boost with weighting function QMCBQ

Classifier Assignment

Make classifier assignment explicit using function

weight of strong classifier on sample

is updated at each round of boosting.

Here: K-component GMM in d-dim eigenspace, k-th mode is area of expertise of

Joint Boosting and Clustering

MCBoost MCBQ

Input: Data set , set of weak learnersOutput: Strong classifiers

for t=1,…,T // boosting roundsfor k=1,…,K // strong classifiers

Find weak learners and their weightsUpdate sample weights

endend

MCBQ Algorithm

Update sample weightsUpdate weighting function

Init with GMMInit weights to values of

, weighting function

MCBQ for Object TrackingPrinciple: 1. (Short) supervised training phase

2. On-line updates

Online Boosting

one sample

Init importance

Estimate errors

Select best weak classifier

Update weight

Estimate importance

Current strong classifier

[Oza, Russel 01, Grabner, Bischof 06]

Global classifier pool

Estimate errors

Select best weak classifier

Update weight

Estimate errors

Select best weak classifier

Update weight

Estimate importance

Online MCBQClassifiers

Sample weight distribution

Selector Selector Selector

Update

Selector Selector Selector

Select weak classifiers, add to

Update weights, re-normalize

Results

Improved Pose Expertise

MCBoost

MCBQ

Multi-pose Tracking with MCBQ

Tracking Experiments

Tracking “Cube” sequence

MCBQMILTrack SemiBoost

Tracking Experiments

Tracking error

Summary

Tracking: Build appearance model, then update online

No detector is required, i.e. not object specific.Handles rapid appearance changes.Simultaneous pose estimation and tracking is possible.

K is currently set by hand.Incorrect adaptation may still occur.

Extension of MCBoost to online settingExtension of MIL to multi-class

Thank you