CWITrack: Transformer tracking via local-global cross-window interaction.

Abstract

Currently, popular trackers using Transformer as the backbone have achieved excellent results. Although Transformers can capture long-range dependencies in sequence data, they perform poorly in learning local information, and transformers typically rely heavily on computational costs to achieve high performance. Existing window-based attention mechanisms, such as the shifted window strategy, achieve cross-window interaction only implicitly through window displacement, requiring multiple stacked layers for effective global modeling. To address these issues, this work designs a Cross-Window Interaction block that includes window attention and window interaction attention, which facilitates local and global information exchange within and among windows in a sequential manner. In addition, a window token embedding has been introduced, which serves as a compact learnable global descriptor for each window, enabling explicit cross-window attention among all windows within a single block with negligible computational cost, thereby achieving long-range information interaction across multiple windows. Upon this, we propose a novel tracker, referred as CWITrack, which can learn local information while focusing on global information, achieving precise and real-time tracking performance. The CWITrack uses an end-to-end trainable manner and has been evaluated on six challenging tracking benchmarks, including UAV123, GOT-10k, TNL2K, LaSOT, TrackingNet and NfS. It achieves superior tracking performance on these benchmarks. Especially, it obtains an AUC of 68.0% on UAV123 and an AO of 72.6% on GOT-10K. The code is available at https://github.com/shafreeli/CWITrack.git.