Bidirectional Consistency Constrained Template Update Learning for Siamese Trackers

Abstract

This paper presents an online template update method with bidirectional consistency constraint for Siamese trackers. Due to continuously applying cross-correlation mechanism between template and the search region, the performance of Siamese trackers highly relies on the fidelity of template. Therefore, besides standard linear update, learning the template update methods attract attention. Inspired by this, in this paper we adopt a learning to update model called UpdateNet as our baseline. Different from it, we further propose a novel bi-directional consistency loss as a constraint to learn the template update more smoothly and stably. Our method considers both forward and backward information for each medium frame, thus introducing a multi-stage bidirectional simulated tracking training mechanism. We apply our model to a Siamese tracker, SiamRPN and demonstrate the effectiveness and robustness of our proposed method compared with traditional UpdateNet in the Large-scale Single Object Tracking (LaSOT) dataset.