Online visual tracking via background-aware Siamese networks

With the rapid development of Siamese network based trackers, a set of related methods have produced considerable performance improvement. However, the tracking results are often disturbed due to the background noise from the template image and background distractor objects from the search image. In this paper, we present an elegant background-aware Siamese tracker for online single object visual tracking. Specifically, a new basic tracking framework is firstly proposed to implement the target localization, bounding box regression, and IoU prediction with offline multi-task learning. During the online tracking stage, we design a novel background-aware tracker with two strategies. Firstly, a spatial mask is introduced to reduce the impacts of background noise from the template image. Secondly, we predict a background-aware salient map to discover and suppress the distractor features in the search image. To validate the effectiveness, we conduct extensive experiments and exhaustive comparisons on OTB2013, OTB2015, VOT2019, UAV123, and GOT10k tracking datasets. Experimental results demonstrate that the proposed tracker, dubbed BaSiamIoU, can achieve state-of-the-art performance while running over 50 FPS.