Tracking your detector performance: How to grow an effective training set in tracking-by-detection methods

In many tracking-by-detection approaches, a self-learning strategy is adopted to augment the training set with new positive and negative instances, and to refine the classifier weights. Previous works focus mainly on the learning algorithm and assume the detector is never wrong while classifying samples at the current frame; the most confident sample is chosen as the target, and the training set is augmented with samples selected in its surrounding area. A wrong choice of such samples may degrade the classifier parameters and cause drifting during tracking. In this paper, the focus is on how samples are chosen while retraining the classifier. A particle filtering framework is used to infer what sample set to add to the training set until some evidence about its correctness becomes available. In preliminary experiments, a simple learning algorithm together with the proposed method to build the training set outperforms our baseline tracking-by-detection algorithm.