Anti-distractors: two-branch siamese tracker with both static and dynamic filters for object tracking

Abstract

Visual Object Tracking is a very challenging task because of the large appearance variance caused by illumination, deformation, and motion. Siamese network-based trackers, which select target through a matching function, are widely used for visual object tracking. The trackers are capable of robustly recognizing the target with appearance variance. However, while the filter template is a crucial part of such methods, most of them did not update the filter template effectively, and have shown limited discriminative ability between target and similar semantic objects (distractors). In order to tackle the challenge of distractors, we added a dynamic filter branch on the traditional siamese network. Under the condition that multipeaks are detected on the static response map, the tracker will redetect target with dynamic branch and the final target location will be determined by the combined result of the dynamic filter branch and static filter branch. Subsequently the sample library with hard negative mining strategy is updated and the dynamic filter kernel is restrained online. With the fusion of two branches, the tracker can distinguish the true target from similar objects. Meanwhile, we conduct extensive experiments and empirical evaluations on two popular datasets: Visdrone and UAV123. Our tracker achieves an AUC of 58% on Visdrone dataset and an AUC of 60.7% on UAV123 dataset.