CSTrack: A Comprehensive and Concise Vision Transformer Tracker

The attention mechanism has been widely applied in visual tracking tasks due to its remarkable ability to capture global dependencies. However, there are two issues in previous attention-based methods: redundancy of template information and inappropriate utilization of information streams. In this work, we propose a spatial positioning attention mechanism that addresses these issues by selective template feature enhancement and elimination of redundant information streams, respectively, significantly improving tracking accuracy and speed. Furthermore, previous trackers fail to focus on channels containing crucial target information within the template features and search region features. To tackle this, we introduce a channel focus attention mechanism to perform channel weight rescaling, which allows the tracker to concentrate on those target-related channels, improving its localization capability. Extensive experiments on four well-known datasets, GOT-10k, LaSOT, TrackingNet, and TNL2K, show that CSTrack outperforms all previous state-of-the-art trackers, running at over 70 FPS.