Research map / Label-efficient and robust learning / Recognition and long-tail
Shallow-Deep Collaborative Learning 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 SDCL, a transformer-based unsupervised visible-infrared ReID framework that collaboratively uses shallow and deep features via collaborative neighbor learning and collaborative ranking association to refine cross-modality labels.
Problem
Unsupervised visible-infrared person re-identification (US-VI-ReID) has no cross-modality identity labels, and prior methods rely only on deep final-layer features, which are unreliable under intra-modality variation and inter-modality discrepancy, causing negative optimization.
Approach
Exploit shallow modality-specific features alongside deep features: neighbor selection and label association are constrained by shallow-deep agreement (homogeneous and heterogeneous), with ranking consistency between shallow and deep similarities used to refine cross-modality pseudo labels.
Key results
- SDCL reaches 64.49% rank-1 on SYSU-MM01 all-search, about 1% above the prior best unsupervised method GUR.
- On RegDB visible-to-infrared, SDCL achieves 86.91% rank-1, about 10 points above GUR.
- Ablation: adding shallow contrastive learning raises rank-1 by 4-7 points over the deep-only baseline on SYSU-MM01 and RegDB; full SDCL is best in all settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add shallow contrastive losses (ablation) | Baseline augmented dual-contrastive learning · SYSU-MM01 all search | rank-1 accuracy: 49.55 -> 53.21 | Helped |
| add shallow contrastive losses (ablation) | Baseline augmented dual-contrastive learning · RegDB visible to infrared | rank-1 accuracy: 63.16 -> 69.97 | Helped |
| add Collaborative Neighbor Learning (CNL) (ablation) | SDCL without CNL or CRA · SYSU-MM01 all search | rank-1 accuracy: 53.21 -> 55.27 | Helped |
| add Collaborative Neighbor Learning (CNL) (ablation) | SDCL without CNL or CRA · RegDB visible to infrared | rank-1 accuracy: 69.97 -> 76.60 | Helped |
| add Collaborative Ranking Association (CRA) (ablation) | SDCL with shallow contrastive learning and CNL · SYSU-MM01 all search | rank-1 accuracy: 59.43 -> 64.49 | Helped |
| add Collaborative Ranking Association (CRA) (ablation) | SDCL with shallow contrastive learning and CNL · RegDB visible to infrared | rank-1 accuracy: 80.53 -> 86.91 | Helped |
| use SDCL instead of GUR | Unsupervised visible-infrared person ReID · SYSU-MM01 all search | rank-1 accuracy: 63.51 -> 64.49 | Helped |
| use SDCL instead of DOTLA | Unsupervised visible-infrared person ReID · RegDB visible to infrared | rank-1 accuracy: 85.63 -> 86.91 | Helped |
Surprising findings
- The unsupervised SDCL outperforms several supervised methods (AGW, MSO, DDAG) on the reported benchmarks.
Limitations
- Relies on DBSCAN clustering for pseudo labels, so label noise from clustering persists.
- Cross-modality label refinement is applied only at the start of later epochs (CRA in last 20 of 50 epochs), which is an offline process.
- Hyperparameters (gamma, N, K, lambda_s) are tuned per dataset.
Experimental setup
Datasets: SYSU-MM01 RegDB
Benchmarks: SYSU-MM01 RegDB
Metrics: rank-1 mAP mINP
Baselines: ADCA GUR DPIS CHCR PGM DOTLA MBCCM DDAG AGW MCLNet
Builds on: TransReID augmented dual-contrastive learning (ADCA) DBSCAN MMT
Closest papers by meaning
- Implicit Discriminative Knowledge Learning for Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.91)
- Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.90)
- TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.90)
- Empowering Visible-Infrared Person Re-Identification with Large Foundation Models (NeurIPS 2024 · similarity 0.90)
- High-Order Structure Based Middle-Feature Learning for Visible-Infrared Person Re-identification (AAAI 2024 · similarity 0.90)
- Relieving Universal Label Noise for Unsupervised Visible-Infrared Person Re-Identification by Inferring from Neighbors (AAAI 2025 · 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.