Research map / Label-efficient and robust learning / Recognition and long-tail
Relieving Universal Label Noise for Unsupervised Visible-Infrared Person Re-Identification by Inferring from Neighbors
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 N-ULC and N-DW, neighbor-based soft pseudo-label calibration and sample weighting added to the PGM framework, to reduce within- and cross-modality label noise in unsupervised visible-infrared person ReID.
Problem
Unsupervised visible-infrared person re-identification (USL-VI-ReID) suffers from label noise within and across modalities due to suboptimal clustering and modality discrepancy, since one-hot pseudo labels cannot represent identities split across clusters.
Approach
Replace one-hot pseudo labels, in both homogeneous and heterogeneous spaces, with soft labels built from the cluster-membership statistics of k-nearest neighbors (N-ULC), and down-weight samples whose neighbors disagree with their pseudo label (N-DW).
Key results
- On SYSU-MM01 all-search, the method reaches 61.81% rank-1, 58.92% mAP and 45.01% mINP, above the best listed USL-VI-ReID baseline GUR (60.95 / 56.99 / 41.85).
- On RegDB infrared-to-visible, the method reaches 88.17% rank-1 and 81.11% mAP, versus 82.91% rank-1 for DOTLA.
- Ablation on SYSU-MM01 all-search: baseline 54.14% rank-1 rises to 55.87 with N-ULC alone and 58.54 with N-DW alone, and 61.81 with both.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add N-ULC to the PGM baseline (ablation) | PGM · SYSU-MM01 Indoor Search | rank-1: 58.80 -> 61.14 | Helped |
| Add N-DW to the PGM baseline (ablation) | PGM · SYSU-MM01 Indoor Search | rank-1: 58.80 -> 62.04 | Helped |
| Add N-ULC and N-DW to the PGM baseline (ablation) | PGM · SYSU-MM01 Indoor Search | rank-1: 58.80 -> 67.04 | Helped |
| Add N-ULC to the PGM baseline (ablation) | PGM · SYSU-MM01 All Search | rank-1: 54.14 -> 55.87 | Helped |
| Add N-DW to the PGM baseline (ablation) | PGM · SYSU-MM01 All Search | rank-1: 54.14 -> 58.54 | Helped |
| Add N-ULC and N-DW to the PGM baseline (ablation) | PGM · SYSU-MM01 All Search | rank-1: 54.14 -> 61.81 | Helped |
| Add N-ULC to the PGM baseline (ablation) | PGM · RegDB Infrared-to-Visible | rank-1: 77.44 -> 79.29 | Helped |
| Add N-DW to the PGM baseline (ablation) | PGM · RegDB Infrared-to-Visible | rank-1: 77.44 -> 87.82 | Helped |
| Add N-ULC and N-DW to the PGM baseline (ablation) | PGM · RegDB Infrared-to-Visible | rank-1: 77.44 -> 88.17 | Helped |
| Add N-ULC to the PGM baseline (ablation) | PGM · RegDB Visible-to-Infrared | rank-1: 77.92 -> 80.60 | Helped |
| Add N-DW to the PGM baseline (ablation) | PGM · RegDB Visible-to-Infrared | rank-1: 77.92 -> 87.88 | Helped |
| Add N-ULC and N-DW to the PGM baseline (ablation) | PGM · RegDB Visible-to-Infrared | rank-1: 77.92 -> 88.75 | Helped |
Surprising findings
- On RegDB the N-DW module alone already gives most of the gain, and adding N-ULC gives only limited extra improvement, which the authors attribute to overlapping functionality.
- The method is reported to beat the USL-VI-ReID baselines on some RegDB columns while the PGM/ DOTLA baselines show comparable mAP on others, so improvements are not uniform across all settings.
Limitations
- Gains on RegDB are limited after N-DW because the two modules overlap in function.
- Evaluated only on SYSU-MM01 and RegDB, both of which are small in identity count and camera setup.
Experimental setup
Datasets: SYSU-MM01 RegDB
Benchmarks: SYSU-MM01 RegDB
Metrics: mAP CMC rank-1 mINP
Baselines: AGW CA MAUM DFLN-ViT PartMix MUN SAAI SPCL MMT CCL ICE PPLR
Builds on: Progressive Graph Matching DBSCAN Random Channel Augmentation
Closest papers by meaning
- Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.91)
- TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.90)
- Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.89)
- Implicit Discriminative Knowledge Learning for Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.87)
- High-Order Structure Based Middle-Feature Learning for Visible-Infrared Person Re-identification (AAAI 2024 · similarity 0.87)
- Empowering Visible-Infrared Person Re-Identification with Large Foundation Models (NeurIPS 2024 · similarity 0.86)
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.