Research map / Label-efficient and robust learning / Recognition and long-tail
TokenMatcher: Diverse Tokens Matching 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 TokenMatcher, which uses multiple class tokens in a ViT (diverse tokens matching), camera-aware neighbor learning, and homogeneous fusion to improve cross-modality pseudo-label matching for unsupervised visible-infrared person re-ID.
Problem
Unsupervised visible-infrared person re-identification must generate reliable cross-modal pseudo-labels without annotations, but single global feature matching is hurt by intra- and inter-modality discrepancies, producing incomplete and unreliable clusters.
Approach
Replace the single global feature with multiple class tokens (diverse embedding spaces) in a transformer, sum per-token cross-modality similarities for matching, refine intra/inter-modality neighbor sets using camera information, and fuse split clusters of the same identity.
Key results
- On RegDB visible-to-infrared, achieves 92.96% rank-1 versus 86.91% for the prior best unsupervised method SDCL, a gain of about 6 points.
- On SYSU-MM01 all-search, achieves 65.07% rank-1 and 62.79% mAP, about 0.6 points rank-1 above SDCL.
- Diverse Tokens Matching alone raises SYSU-MM01 all-search rank-1 by about 11 points and RegDB visible-to-infrared rank-1 by about 38 points over baseline in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TokenMatcher instead of SDCL | Unsupervised VI-ReID · SYSU-MM01 All Search | rank-1 accuracy: 64.49 -> 65.07 | Helped |
| Use TokenMatcher instead of SDCL | Unsupervised VI-ReID · SYSU-MM01 Indoor Search | rank-1 accuracy: 71.37 -> 68.97 | Hurt |
| Use TokenMatcher instead of SDCL | Unsupervised VI-ReID · RegDB Visible to Infrared | rank-1 accuracy: 86.91 -> 92.96 | Helped |
| Use TokenMatcher instead of SDCL | Unsupervised VI-ReID · RegDB Infrared to Visible | rank-1 accuracy: 85.76 -> 91.82 | Helped |
| Add DTM (ablation) | TokenMatcher baseline · SYSU-MM01 All Search | rank-1 accuracy: 51.27 -> 53.54 | Helped |
| Add DTNL (ablation) | TokenMatcher baseline · SYSU-MM01 All Search | rank-1 accuracy: 51.27 -> 54.35 | Helped |
| Add HF (ablation) | TokenMatcher baseline · SYSU-MM01 All Search | rank-1 accuracy: 51.27 -> 62.16 | Helped |
| Add DTM and DTNL (ablation) | TokenMatcher baseline · SYSU-MM01 All Search | rank-1 accuracy: 51.27 -> 64.52 | Helped |
| Add DTM, DTNL, and HF (ablation) | TokenMatcher baseline · SYSU-MM01 All Search | rank-1 accuracy: 51.27 -> 65.07 | Helped |
| Add DTM (ablation) | TokenMatcher baseline · SYSU-MM01 Indoor Search | rank-1 accuracy | Helped |
| Add DTNL (ablation) | TokenMatcher baseline · RegDB Visible to Infrared | rank-1 accuracy | Helped |
| Add HF (ablation) | TokenMatcher baseline · RegDB Infrared to Visible | rank-1 accuracy | Helped |
Surprising findings
- Gains on SYSU-MM01 over the prior unsupervised SOTA are small (about 0.6 rank-1), while RegDB gains are large.
Limitations
- Performance remains far below fully supervised VI-ReID methods due to lack of annotations.
- Number of class tokens N and matching hyperparameters (lambda, beta weights) are dataset-specific (N=4 for SYSU-MM01, 6 for RegDB).
Experimental setup
Datasets: SYSU-MM01 RegDB
Benchmarks: SYSU-MM01 RegDB
Metrics: rank-1 mAP mINP
Baselines: ADCA PGM MBCCM DOTLA DCCL GUR SDCL OTLA
Builds on: DC-Former ADCA PGM CALR SDCL
Closest papers by meaning
- Modality-Aware Bias Mitigation and Invariance Learning for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2026 · similarity 0.92)
- Shallow-Deep Collaborative Learning for Unsupervised Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.90)
- Relieving Universal Label Noise for Unsupervised Visible-Infrared Person Re-Identification by Inferring from Neighbors (AAAI 2025 · similarity 0.90)
- Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.90)
- Implicit Discriminative Knowledge Learning for Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.89)
- Empowering Visible-Infrared Person Re-Identification with Large Foundation Models (NeurIPS 2024 · similarity 0.89)
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.