Reproduce “UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hao Wu, Xudong Wang, Jialiang Zhang, Junlong Tong, Xinghao Chen, Junyan Lin, Yunpu Ma, Xiaoyu Shen
The first unified token pruning framework for one-stream Transformer-based visual tracking, UTPTrack jointly compresses search, dynamic, and static template tokens using an attention-guided approach. It achieves up to a 67.5% vision token reduction and 31.3% MAC reduction while maintaining or improving accuracy across 10 RGB-based and multimodal tracking benchmarks.
Key insights
- Jointly modeling and compressing tokens across all visual components (search region, static template, dynamic template) provides superior accuracy-efficiency trade-offs compared to fragmented, component-isolated pruning.
- Reusing the Transformer's intrinsic attention weights to derive token importance scores offers an efficient pruning mechanism that avoids the need for auxiliary prediction modules.
- Integrating spatial priors (Token Type-Aware pruning) for static template tokens and semantic cues (Text-Guided pruning) into the pruning process significantly enhances robustness and can even improve baseline accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.