Research map / Vision-language and multimodal / Cross-modal learning
Noisy-Correspondence Learning for Text-to-Image 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 RDE, a dual-embedding CLIP-based text-to-image person re-ID method that handles noisy image-text correspondences via a GMM-based consensus division of clean pairs and a log-exponential upper-bound triplet loss (TAL).
Problem
Text-to-image person re-identification assumes image-text training pairs are correctly aligned, but real-world annotation errors and low-quality images create noisy correspondences that degrade and overfit models.
Approach
Combines a basic global embedding and a token selection embedding, uses a two-component GMM on per-sample losses from both to take the consensus clean set (relabeling the rest as noisy or uncertain), and replaces the hardest-negative triplet ranking loss with a log-exponential upper bound over all negatives (TAL) to avoid model collapse under noise.
Key results
- On CUHK-PEDES with 50% synthetic noise, RDE reaches 71.33% Rank-1 (best checkpoint), +6.11% over the best baseline DECL.
- Without synthetic noise, RDE improves Rank-1 over IRRA by +2.56% on CUHK-PEDES, +4.22% on ICFG-PEDES and +5.15% on RSTPReid.
- Ablation on CUHK-PEDES at 50% noise: TRL collapses to 6.40% Rank-1 while TAL reaches 71.33%, and TRL-S reaches 67.38%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RDE instead of IRRA | TIReID with CLIP ViT-B/16 · CUHK-PEDES (0% noise) | Rank-1 accuracy: 73.38 -> 75.94 | Helped |
| Use RDE instead of IRRA | TIReID with CLIP ViT-B/16 · ICFG-PEDES (0% noise) | Rank-1 accuracy: 63.46 -> 67.68 | Helped |
| Use RDE instead of IRRA | TIReID with CLIP ViT-B/16 · RSTPReid (0% noise) | Rank-1 accuracy: 60.20 -> 65.35 | Helped |
| Use RDE instead of DECL | TIReID with CLIP ViT-B/16 · CUHK-PEDES (50% noise) | Rank-1 accuracy: 65.22 -> 71.33 | Helped |
| Use RDE instead of DECL | TIReID with CLIP ViT-B/16 · ICFG-PEDES (50% noise) | Rank-1 accuracy: 57.50 -> 63.76 | Helped |
| Use RDE instead of DECL | TIReID with CLIP ViT-B/16 · RSTPReid (50% noise) | Rank-1 accuracy: 56.75 -> 62.85 | Helped |
| Replace TAL with TRL (ablation) | RDE · CUHK-PEDES (50% noise) | Rank-1 accuracy: 71.33 -> 6.40 | Hurt |
| Replace TAL with TRL-S (ablation) | RDE · CUHK-PEDES (50% noise) | Rank-1 accuracy: 71.33 -> 67.38 | Hurt |
| Replace TAL with SDM (ablation) | RDE · CUHK-PEDES (50% noise) | Rank-1 accuracy: 71.33 -> 69.33 | Hurt |
| Remove TSE (ablation) | RDE · CUHK-PEDES (50% noise) | Rank-1 accuracy: 71.33 -> 70.70 | Hurt |
| Remove BGE (ablation) | RDE · CUHK-PEDES (50% noise) | Rank-1 accuracy: 71.33 -> 69.07 | Hurt |
| Remove CCD (ablation) | RDE · CUHK-PEDES (50% noise) | Rank-1 accuracy: 71.33 -> 63.11 | Hurt |
Surprising findings
- Standard triplet ranking loss with hardest negatives collapses under noisy correspondence (Rank-1 of 6.40% at 50% noise on CUHK-PEDES), with similarities all converging toward 1.
Limitations
- Relies on synthetic random text shuffling for noise, which may not match real-world noise patterns.
- Performance at very small tau causes training failure; hyperparameters (m, tau, R) require tuning.
Experimental setup
Datasets: CUHK-PEDES ICFG-PEDES RSTPReid
Benchmarks: CUHK-PEDES ICFG-PEDES RSTPReid
Metrics: Rank-1 Rank-5 Rank-10 mAP mINP
Baselines: SSAN IVT IRRA DECL CLIP-C CFine TRL TRL-S SDM loss
Builds on: IRRA CLIP Triplet Ranking loss DECL memorization effect sample selection
Closest papers by meaning
- Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.89)
- Unifying Multi-Modal Uncertainty Modeling and Semantic Alignment for Text-to-Image Person Re-identification (AAAI 2024 · similarity 0.89)
- Learning With Twin Noisy Labels for Visible-Infrared Person Re-Identification (CVPR 2022 · similarity 0.89)
- KPDM: Key Phrase Dynamic Masking for Robust Text-to-Image Person Retrieval (AAAI 2026 · similarity 0.89)
- Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic (ICLR 2024 · similarity 0.86)
- Modeling Thousands of Human Annotators for Generalizable Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.85)
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.