Online Multiple Classifier Boosting for Object Tracking Tae-Kyun Kim1 Thomas Woodley1 Björn Stenger2 Roberto Cipolla1 1Dept. of Engineering, University of Cambridge 2Computer Vision Group, Toshiba Research Europe The Task: Object Tracking Example sequence 1 Example sequence 2 Target appearance changes due to changes in - pose - illumination - object deformation Learning Multi-Modal Representations Positive examples Negative examples - Multi-view face detection [Rowley et al. 98, Schneiderman et al. 00, Jones Viola 03] - Multi-category detection, Sharing features [Torralba et al. 04] Joint Clustering and Training [Kim and Cipolla 08, Babenko et al. 08] Positive examples Feature pool Negative examples Face cluster 1 K-means clustering Face cluster 2 MCBoost: Multiple Strong Classifier Boosting [Kim and Cipolla 08, Babenko et al. 08] Given: Set of n training samples with labels number of strong classifiers Learn strong classifiers: Map to probabilities with sigmoid function Combine classifier output with “Noisy OR” function MCBoost (continued) • 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: Toy Example 1 Input data MCBoost result (K=3) Toy Example 2 Standard AdaBoost MCBoost [Kim and Cipolla 08] MC Boost with weighting function Q MCBQ 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 MCBQ Algorithm Input: Data set , set of weak learners , weighting function Output: Strong classifiers Init with GMM Init weights to values of for t=1,…,T // boosting rounds for k=1,…,K // strong classifiers Find weak learners and their weights Update sample weights Update weighting function end end MCBQ for Object Tracking Principle: 1. (Short) supervised training phase 2. On-line updates Online Boosting [Oza, Russel 01, Grabner, Bischof 06] Global classifier pool one sample Init importance Estimate errors Estimate errors Select best weak classifier Update weight Estimate importance Select best weak classifier Update weight Current strong classifier Estimate errors Estimate importance Select best weak classifier Update weight Online MCBQ Classifiers Sample weight distribution Select weak classifiers, add to Selector Selector Selector Update Update weights, re-normalize Selector Selector Selector Results Improved Pose Expertise MCBoost MCBQ Multi-pose Tracking with MCBQ Tracking Experiments Tracking “Cube” sequence MILTrack SemiBoost MCBQ Tracking Experiments Tracking error Summary Extension of MCBoost to online setting Extension of MIL to multi-class 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. Thank you