Spatiotemporal key region transformer for visual tracking

Visual tracking is an important field of computer vision research. Although transformer-based trackers have achieved remarkable performance, the transformer structure is globally computationally inefficient, it does not screen important patches, and it cannot focus on key target regions. At the same time, temporal motion features are easily overlooked. To solve these problems, this paper proposes a new method, SKRT, that removes the CNN structure and directly uses a transformer as the backbone network to extract multiframe video features. Then, these feature maps are mixed and superimposed to obtain spatiotemporal information. To focus on important parts efficiently, we use key region extraction to obtain a small set of template and search feature map patches and reinput them into the transformer as a cross-correlation computation. Finally, we predict the position of a tracking object through center-corner prediction. To demonstrate the effectiveness of our method, we conduct experiments on challenging benchmark datasets (GOT-10K, TrackingNet, VOT2018, OTB100, LaSOT), and the results show that SKRT is competitive with other state-of-the-art methods.