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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformers excel at long-range dependencies but struggle with local information and high computational costs.
• Existing window-based attention mechanisms offer limited cross-window interaction, requiring deep networks for global context.

Purpose of the Study:

• To develop an efficient visual tracker that balances local and global information processing.
• To overcome the limitations of current Transformer-based trackers in terms of local feature learning and computational expense.

Main Methods:

• Introduced a Cross-Window Interaction block with window attention and interaction attention for sequential local and global information exchange.
• Developed a window token embedding for explicit, low-cost cross-window attention, enabling efficient long-range information interaction.

Main Results:

• CWITrack demonstrates superior performance on six challenging benchmarks (UAV123, GOT-10k, TNL2K, LaSOT, TrackingNet, NfS).
• Achieved an AUC of 68.0% on UAV123 and an AO of 72.6% on GOT-10K.
• The tracker is end-to-end trainable and achieves precise, real-time tracking.

Conclusions:

• CWITrack effectively integrates local and global information processing within a single block.
• The proposed methods significantly improve tracking accuracy and efficiency compared to existing approaches.