MSTrack: Visual Tracking with Multi-scale Attention
Chunlin Song, Yu Yao, Jianhui Guo, Lunbo Li · 2024
The attention mechanism has been widely applied to various computer vision tasks due to it excels at capturing global feature dependencies. However, for visual tracking tasks, conventional attention mechanisms model feature dependencies based on feature maps of only one size, which hinders the ability of the tracker to effectively handle target scale variations. To address this issue, we propose a novel multi-scale attention mechanism that captures global dependencies between the template and the search region from feature maps of various sizes, enhancing the sensitivity of the tracker to scale variations. Furthermore, our attention residual operation employs an attention prior to guide the modeling of small-size feature dependencies, effectively prioritizing the focus on primary target information. Extensive experimental results on widely-used tracking datasets GOT-10k, LaSOT, TNL2K, and TrackingNet, demonstrate that our proposed one-stream tracker MSTrack outperforms all previous state-of-the-art trackers, running at 100.5 FPS.