Object Tracking with Multiple Dynamic Templates Updating

Abstract

AbstractMost existing Siamese visual trackers see the object tracking task as similarity learning between a search image and a single template image. Utilizing only one template leads to the negligence of the rich semantic information in other frames. Meanwhile, those Siamese trackers with temporal context exploitation either incorporate specially designed non-generic modules or include online-learning parts which compromise real-time performance. In this paper, we propose a novel model architecture incorporating multiple dynamic templates in a Siamese visual tracker to maximize temporal information utilization. To attain a favorable appearance representation from these templates, we propose an online dynamic template pool updater that leverages the frames with dissimilar appearances. Furthermore, we design a new hard positive sampling strategy to train the tracker with dissimilar templates. With the proposed methods, a Siamese tracker can be straightforwardly transformed and trained to benefit from the temporal correlations among frames. Comprehensive experiments on various tracking datasets show positive results and prove the effectiveness of the proposed methods.KeywordsObject trackingDeep learningSiamese networksDynamic templatesHard positive sampling