Research map / Label-efficient and robust learning / Recognition and long-tail
Identity-Clothing Similarity Modeling for Unsupervised Clothing Change 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 ICSM, a three-stage unsupervised clothing-change ReID framework that uses identity/clothing cluster-level CLIP text features to connect clothing-change positive clusters and adaptively optimizes only reliable pairs.
Problem
Unsupervised clothing-change person re-identification struggles because clustering fails to assign the same pseudo-label to images of the same person wearing different clothing, and standard contrastive/classification losses push these clothing-change positive pairs apart.
Approach
Clusters compact clothing-aware pseudo-labels, then represents each cluster with learnable-text CLIP features for identity (from body-only images) and clothing (from clothing-only images), uses their similarities to find clothing-change positive clusters and reliable/hard negatives, and optimizes adaptive ID and hard-contrastive losses restricted to those reliable sets plus text-semantic contrasti
Key results
- On PRCC(CC), ICSM reaches 57.8 mAP / 47.9 Rank-1, outperforming the best prior unsupervised UCC-ReID method CICL (55.3 / 45.0).
- On VC-Clothes(CC), ICSM achieves 73.5 mAP / 75.1 Rank-1 versus CICL's 69.3 / 72.8.
- Ablation shows ISE alone raises PRCC(CC) mAP from 43.9 (baseline) to 53.6; adding both text-semantic losses reaches 57.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · PRCC(CC) | mAP: 55.3 -> 57.8 | Helped |
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · PRCC(SC) | mAP: 96.7 -> 97.5 | Helped |
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · VC-Clothes(CC) | mAP: 69.3 -> 73.5 | Helped |
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · VC-Clothes(G) | mAP: 76.7 -> 78.7 | Helped |
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · LTCC(CC) | mAP: 13.0 -> 13.5 | Helped |
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · LTCC(G) | mAP: 29.6 -> 30.1 | Helped |
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · Celeb-reID(CC) | mAP: 6.7 -> 7.0 | Helped |
| Apply ICSM instead of CICL | Unsupervised clothing-change person re-identification · Celeb-reID-light(CC) | mAP: 7.8 -> 7.7 | Hurt |
| Add ISE and use adaptive losses instead of the stage-I losses (ablation) | M1 clothing-change ReID baseline · PRCC(CC) | mAP: 43.9 -> 53.6 | Helped |
| Add ISE and use adaptive losses instead of the stage-I losses (ablation) | M1 clothing-change ReID baseline · VC-Clothes(CC) | mAP: 23.6 -> 57.6 | Helped |
| Add identity-oriented semantic contrastive loss (ablation) | M2 with ISE and adaptive losses · VC-Clothes(CC) | mAP: 57.6 -> 68.6 | Helped |
| Add clothing-invariant semantic contrastive loss (ablation) | M2 with ISE and adaptive losses · VC-Clothes(CC) | mAP: 57.6 -> 63.8 | Helped |
Surprising findings
- Clustering clothing-occluded images performs worse than clustering original images, because clothing cues help form compact same-clothing clusters.
Limitations
- Relies on a human parsing model to produce identity-related and clothing-related images, adding dependency on parser quality.
- Three-stage pipeline with many hyperparameters (gamma_a, gamma_c, gamma_b, lambda_is, lambda_cs) tuned on test-related sensitivity analyses.
Experimental setup
Datasets: PRCC VC-Clothes LTCC Celeb-reID Celeb-reID-light
Benchmarks: PRCC VC-Clothes LTCC Celeb-reID Celeb-reID-light
Metrics: mAP CMC Rank-1
Baselines: ICE CC CACL PPLR RTMem Purification DCCC SiCL CICL RCSANet GI-ReID CAL
Builds on: DBSCAN CLIP ICE SiCL CICL
Closest papers by meaning
- DIFFER: Disentangling Identity Features via Semantic Cues for Clothes-Changing Person Re-ID (CVPR 2025 · similarity 0.91)
- Graph-Enhanced Attribute-Aware Modeling for Cloth-Changing Person Re-Identification (NeurIPS 2026 · similarity 0.91)
- Dual-stream Relation-modeling Disentanglement for Cloth-Changing Person Re-Identification (AAAI 2026 · similarity 0.90)
- COCAS: A Large-Scale Clothes Changing Person Dataset for Re-Identification (CVPR 2020 · similarity 0.89)
- DKC: Differentiated Knowledge Consolidation for Cloth-Hybrid Lifelong Person Re-identification (CVPR 2025 · similarity 0.87)
- Instruct-ReID: A Multi-purpose Person Re-identification Task with Instructions (CVPR 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.