Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PCLHD, a progressive contrastive learning method for unsupervised visible-infrared person re-ID that adds hard prototypes (farthest sample from centroid) and dynamic prototypes (random cluster samples) on top of centroid-based contrastive learning.
Problem
Unsupervised visible-infrared person re-identification (USVI-ReID) without annotations; existing cluster-center (centroid) prototypes capture only commonality, yielding unreliable pseudo-labels and cross-modality correspondences.
Approach
Replace or complement the single centroid prototype with a hard prototype (the sample farthest from the cluster center, used to capture divergence) and dynamic prototypes (randomly sampled cluster members, used to capture variety), introduced progressively after an initial centroid-only stage so that early unreliable clusters do not degrade.
Key results
- PCLHD reaches 64.4% Rank-1 and 58.7% mAP on SYSU-MM01 all-search, the best among USVI-ReID methods without MMM; PCLHD+MMM reaches 65.9% Rank-1 and 61.8% mAP.
- Ablation on SYSU-MM01 all-search: baseline PGM 56.3% Rank-1; adding HPCL and DPCL together gives 63.7%, and full PCL gives 64.4%.
- On RegDB Visible2Thermal, PCLHD reports 84.3% Rank-1 and 80.7% mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HPCL to the PGM baseline (ablation) | PGM · SYSU-MM01 all-search | Rank-1: 56.3 -> 62.1 | Helped |
| add HPCL to the PGM baseline (ablation) | PGM · SYSU-MM01 all-search | mAP: 51.7 -> 56.8 | Helped |
| add DPCL to the PGM baseline (ablation) | PGM · SYSU-MM01 all-search | Rank-1: 56.3 -> 59.1 | Helped |
| add DPCL to the PGM baseline (ablation) | PGM · SYSU-MM01 all-search | mAP: 51.7 -> 54.4 | Helped |
| add HPCL and DPCL to the PGM baseline (ablation) | PGM · SYSU-MM01 all-search | Rank-1: 56.3 -> 63.7 | Helped |
| add HPCL and DPCL to the PGM baseline (ablation) | PGM · SYSU-MM01 all-search | mAP: 51.7 -> 57.8 | Helped |
| add progressive learning (PCL) to HPCL and DPCL (ablation) | PGM + HPCL + DPCL · SYSU-MM01 all-search | Rank-1: 63.7 -> 64.4 | Helped |
| add progressive learning (PCL) to HPCL and DPCL (ablation) | PGM + HPCL + DPCL · SYSU-MM01 all-search | mAP: 57.8 -> 58.7 | Helped |
| set λ=1, removing DPCL from PCLHD (ablation) | PCLHD · SYSU-MM01 all-search | Rank-1 | Hurt |
| vary hard-prototype sample count k from 1 to 3 (ablation) | PCLHD · SYSU-MM01 all-search | Rank-1 | Hurt |
| combine PCLHD with MMM | PCLHD · SYSU-MM01 all-search | Rank-1: 61.6 -> 65.9 | Helped |
Surprising findings
- Using more than one hard sample (k>1) as the hard prototype degrades performance, suggesting one prototype suffices.
- Setting lambda=1 (no DPCL) causes a significant performance drop.
Limitations
- Relies on DBSCAN for pseudo-labels, which may not scale to extremely large datasets.
- No error bars reported due to computational cost.
- Still a significant gap to fully supervised state-of-the-art results.
Experimental setup
Datasets: SYSU-MM01 RegDB
Benchmarks: SYSU-MM01 RegDB
Metrics: Rank-1 mAP ARI
Baselines: PGM MMM GUR CCLNet MBCCM NGLR ADCA OTLA H2H DDAG AGW DPIS
Builds on: PGM MoCo AGW DBSCAN CAJ
Closest papers by meaning
- Relieving Universal Label Noise for Unsupervised Visible-Infrared Person Re-Identification by Inferring from Neighbors (AAAI 2025 · similarity 0.91)
- Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.90)
- TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.90)
- High-Order Structure Based Middle-Feature Learning for Visible-Infrared Person Re-identification (AAAI 2024 · similarity 0.89)
- Empowering Visible-Infrared Person Re-Identification with Large Foundation Models (NeurIPS 2024 · similarity 0.88)
- Diffusion-based Synthetic Data Generation for Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.