Research map / Vision-language and multimodal / Cross-modal learning
Text-Based Occluded Person Re-identification via Multi-Granularity Contrastive Consistency Learning
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.
Builds occluded versions of three text-based person re-ID benchmarks with an occlusion generator (OGor) and proposes MGCC, a CLIP-based multi-granularity contrastive consistency model with fine-grained token selection.
Problem
Text-based person re-identification fails when gallery images are occluded and inconsistent, since existing T-ReID benchmarks assume full-body, unoccluded visuals and there are no occluded text-based ReID datasets.
Approach
An occlusion generator pastes real occluder instances onto surveillance images to create occluded T-ReID benchmarks, and a CLIP-based model aligns image and text at coarse and fine granularities with attention-based fusion, after pruning uninformative tokens via [CLS] attention scores.
Key results
- MGCC reaches 62.44% R@1 on Occluded-CUHK-PEDES, 59.28% R@1 on Occluded-ICFG-PEDES, and 49.85% R@1 on Occluded-RSTPReid.
- On Occluded-RSTPReid, MGCC improves R@1 by 1.2 points over IRRA (49.85 vs 48.65).
- Token selection with ratios (0.3, 0.4) reduces memory by about 16-17% and inference time by 28-50% with a slight R@1 change.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MGCC instead of IRRA | CLIP-CLIP · Occluded-CUHK-PEDES | R@1: 64.41 -> 62.44 | Hurt |
| Use MGCC instead of IRRA | CLIP-CLIP · Occluded-ICFG-PEDES | R@1: 57.24 -> 59.28 | Helped |
| Use MGCC instead of IRRA | CLIP-CLIP · Occluded-RSTPReid | R@1: 48.65 -> 49.85 | Helped |
| Add fine-granularity parallel contrast SP-W to SI-T (ablation) | MGCC · Occluded-CUHK-PEDES | R@1: 57.33 -> 62.23 | Helped |
| Add fine-granularity parallel contrast SP-W to SI-T (ablation) | MGCC · Occluded-ICFG-PEDES | R@1: 57.88 -> 58.52 | Helped |
| Add fine-granularity parallel contrast SP-W to SI-T (ablation) | MGCC · Occluded-RSTPReid | R@1: 41.30 -> 48.80 | Helped |
| Add inter-granularity cross contrast SP-T to SI-T (ablation) | MGCC · Occluded-CUHK-PEDES | R@1: 57.33 -> 61.92 | Helped |
| Add inter-granularity cross contrast SP-T to SI-T (ablation) | MGCC · Occluded-ICFG-PEDES | R@1: 57.88 -> 58.60 | Helped |
| Add inter-granularity cross contrast SP-T to SI-T (ablation) | MGCC · Occluded-RSTPReid | R@1: 41.30 -> 46.55 | Helped |
| Add fine-granularity token selection using the reported ratios (ablation) | MGCC · Occluded-CUHK-PEDES | R@1: 62.75 -> 62.44 | Hurt |
| Add fine-granularity token selection using the reported ratios (ablation) | MGCC · Occluded-ICFG-PEDES | R@1: 58.80 -> 59.28 | Helped |
| Add fine-granularity token selection using the reported ratios (ablation) | MGCC · Occluded-RSTPReid | R@1: 48.35 -> 49.85 | Helped |
Surprising findings
- On Occluded-CUHK-PEDES, MGCC is slightly below IRRA on R@1 and R@5 (62.44 vs 64.41), though it has the best Rsum.
Limitations
- Occlusions are synthetic, pasted from a fixed library of 15 occluder classes rather than real occlusions.
- MGCC does not beat IRRA on R@1 and R@5 for Occluded-CUHK-PEDES.
Experimental setup
Datasets: CUHK-PEDES ICFG-PEDES RSTPReid
Benchmarks: Occluded-CUHK-PEDES Occluded-ICFG-PEDES Occluded-RSTPReid
Metrics: R@1 R@5 R@10 mAP Rsum
Baselines: DPath CMPM TBPS TIPCB ViTAA NAFS SAF SSAN CFine LGUR IRRA
Builds on: CLIP InfoNCE IRRA Mask R-CNN
Closest papers by meaning
- Unifying Multi-Modal Uncertainty Modeling and Semantic Alignment for Text-to-Image Person Re-identification (AAAI 2024 · similarity 0.88)
- Dynamic Feature Pruning and Consolidation for Occluded Person Re-identification (AAAI 2024 · similarity 0.88)
- Modeling Thousands of Human Annotators for Generalizable Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.85)
- Noisy-Correspondence Learning for Text-to-Image Person Re-identification (CVPR 2024 · similarity 0.85)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.85)
- Automatic Synthetic Data and Fine-grained Adaptive Feature Alignment for Composed Person Retrieval (NeurIPS 2025 · similarity 0.84)
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.