Robust visual tracking with a novel online semi-supervised multiple instance boosting algorithm
Si Chen, Shaozi Li · 2013
Tracking-by-detection is recently formulated as a multiple instance learning (MIL) problem. However, the existing MIL trackers update the weak classifiers with positive labels for all the instances in the positive bag for each frame, which may decrease the tracking performance considerably. In this paper we propose a novel online semi-supervised multiple instance boosting algorithm, termed SemiMILBoost, to achieve robust visual tracking. We employ an effective online updating framework, where the weak classifiers are iteratively updated using the pseudo-labels of all the instances in the positive bag which are predicted by the semi-supervised learning method. Furthermore, a new weighted bag probability function is used to choose the best weak classifiers by introducing the instance weights, and then we minimize the negative bag log likelihood via the functional gradient descent technique. Experimental results demonstrate that our proposed algorithm outperforms the state-of-the-art tracking methods on several challenging video sequences.