Reproduce “SCORE: SubDistribution-aware Collaborative Knowledge Reinforcing for Cloth-Hybrid Lifelong 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

Kunlun Xu, Liangyu Ma, Jiangmeng Li, Xin Tong, Xiaode Liu, Yufei Guo, Jiahuan Zhou

SCORE addresses lifelong person re-identification when training streams alternate between cloth-consistent and cloth-changing datasets, where clothing cues can both help and confuse identity matching. By representing each identity with multiple Gaussian subdistributions and retaining distribution- and instance-level relations without image replay, it improves overall performance over DKC, reaching up to 47.8% mAP and 59.7% Rank-1 while reducing forgetting.

Key insights

  • Intra-identity variation should be modeled as structured information rather than discarded as noise, particularly when clothing changes across learning stages.
  • Preserving distribution spread and relationships among samples provides more retention than storing prototype centers alone; subdistribution modeling supplies the largest ablation gain, with relational losses adding further improvements.
  • SCORE preserves both clothing-relevant and clothing-invariant cues instead of forcing the representation to eliminate clothing information entirely.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.