Reproduce “TS-CAM: Token Semantic Coupled Attention Map for Weakly Supervised Object Localization” 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
Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, Qixiang Ye
TS-CAM introduces a method for Weakly Supervised Object Localization that effectively adapts Visual Transformers to address the persistent 'partial activation' issue in existing approaches. The framework combines semantic-aware feature representations with global spatial attention to accurately localize entire objects, achieving a 71.3% Top-1 Localization Accuracy on CUB-200-2011 and 53.4% on ILSVRC, representing substantial improvements over prior art.
Key insights
- The long-range dependency capabilities of Visual Transformers can fundamentally overcome the 'partial activation' problem prevalent in CNN-based WSOL methods.
- For effective transformer-based WSOL, it is crucial to convert semantic-agnostic attention maps into class-specific representations while retaining global spatial context.
- Coupling re-allocated semantic information from patch tokens with aggregated attention from the class token enables the generation of precise localization maps covering entire object regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.