Reproduce “From Global to Local: Rethinking CLIP Feature Aggregation for Person Re-Identification” 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
Aotian Zheng, Winston Sun, Bahaa Alattar, Vitaly Ablavsky, Jenq-Neng Hwang
SAGA-ReID, developed at the University of Washington, refines CLIP-based person re-identification by introducing a structured anchor-guided aggregation mechanism to overcome the limitations of global pooling. This approach significantly improves robustness against occlusion and background clutter, achieving up to a +10.6 Rank-1 improvement on occluded benchmarks while maintaining strong performance on standard datasets.
Key insights
- The fragility of CLIP-based ReID stems from an "aggregation structure gap," where the global [CLS] token's attention pattern is not optimized for discerning identity-relevant regions from corrupted areas.
- CLIP's text embedding space can serve as a structured optimization prior for feature aggregation, guiding anchors along semantically meaningful directions for more diverse and stable spatial selectivity.
- Reconstructed features, which selectively emphasize stable identity evidence, are complementary to global [CLS] features, with score-level fusion yielding the best performance, especially under occlusion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.