Reproduce “Identity-Seeking Self-Supervised Representation Learning for Generalizable 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
Zhaopeng Dou, Zhongdao Wang, Yali Li, Shengjin Wang
The Identity-seeking Self-supervised Representation learning (ISR) method learns domain-generalizable person re-identification representations directly from large-scale unlabeled internet videos. It achieves state-of-the-art results in domain generalization without human annotation or fine-tuning, outperforming existing supervised and conventional self-supervised approaches.
Key insights
- Learning explicit 'identity discrimination' through cross-frame positive pair mining is more effective for ReID than instance-level discrimination used in generic self-supervised learning.
- Robustly managing noise in self-generated pseudo-labels, especially with a reliability-guided loss and gradient cut-off, is critical for successful self-supervised learning on real-world, uncurated data.
- Large-scale, readily available unlabeled video data, despite its inherent noise, can be a superior resource for training highly generalizable ReID models compared to limited annotated datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.