Research map / Label-efficient and robust learning / Recognition and long-tail
DIFFER: Disentangling Identity Features via Semantic Cues for Clothes-Changing Person Re-ID
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 DIFFER, which uses VLM-generated biometric and non-biometric text descriptions as supervision and a gradient-reversal NBDetach module to disentangle identity features for clothes-changing person re-ID using RGB images only.
Problem
Clothes-changing person re-identification needs identity features that are robust to clothing, pose, and hairstyle changes, but existing methods rely on extra modalities (silhouette, pose, mesh) or discrete attribute labels that cannot capture the variability of non-biometric factors.
Approach
Generate separate biometric and non-biometric textual descriptions with a VLM (CogVLM), encode them as contrastive targets, and use gradient reversal layers on non-biometric projection heads so the image encoder sheds non-biometric information while keeping biometric information, with no extra modality or text at inference.
Key results
- DIFFER improves Rank-1 on LTCC from 54.6% to 58.2% over the ID-loss baseline with EVA02-CLIP-L.
- On PRCC (cloth-changing), DIFFER reaches 68.5% Rank-1 and 64.7% mAP, versus 65.1% and 63.0% for the baseline.
- Top-1 gains of 2.5% on Celeb-reID-light and 1% on CCVID over the baseline; DIFFER outperforms listed prior methods on PRCC and LTCC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add NBDetach to the ID-loss baseline | DIFFER · LTCC, clothes-changing | top-1 accuracy: 54.6 -> 58.2 | Helped |
| Add NBDetach to the ID-loss baseline | DIFFER · LTCC, clothes-changing | mAP: 30.4 -> 31.6 | Helped |
| Add NBDetach to the ID-loss baseline | DIFFER · PRCC, clothes-changing | top-1 accuracy: 65.1 -> 68.5 | Helped |
| Add NBDetach to the ID-loss baseline | DIFFER · PRCC, clothes-changing | mAP: 63.0 -> 64.7 | Helped |
| Add NBDetach to the ID-loss baseline | DIFFER · CCVID, clothes-changing | top-1 accuracy: 86.9 -> 87.9 | Helped |
| Add NBDetach to the ID-loss baseline | DIFFER · CelebReID-Light | top-1 accuracy: 72.7 -> 75.6 | Helped |
| Add biometric contrastive loss to ID loss (ablation) | DIFFER · LTCC, clothes-changing | top-1 accuracy: 54.6 -> 55.4 | Helped |
| Add non-biometric contrastive loss to ID loss (ablation) | DIFFER · LTCC, clothes-changing | top-1 accuracy: 54.6 -> 55.6 | Helped |
| Use both biometric and non-biometric contrastive losses (ablation) | DIFFER · LTCC, clothes-changing | top-1 accuracy: 54.6 -> 58.2 | Helped |
| Use both biometric and non-biometric contrastive losses (ablation) | DIFFER · PRCC, clothes-changing | top-1 accuracy: 65.1 -> 68.5 | Helped |
| Use hair as the non-biometric factor instead of none (ablation) | DIFFER · LTCC | top-1 accuracy: 55.4 -> 56.6 | Helped |
| Use clothing as the non-biometric factor instead of none (ablation) | DIFFER · LTCC | top-1 accuracy: 55.4 -> 56.6 | Helped |
Surprising findings
- Combining multiple non-biometric factors reduced performance when the subspace dimension matched the image feature dimension (1024), with Rank-1 dropping from 57.1% to 54.1%.
Limitations
- Combining multiple non-biometric factors can destabilize the gradient reversal process.
- Retrieval failures remain for occluded individuals and visually similar body shapes and hairstyles.
- Ethical concerns: person re-identification could enable privacy violations.
Experimental setup
Datasets: LTCC PRCC Celeb-reID-light CCVID
Benchmarks: LTCC PRCC CCVID Celeb-reID-light
Metrics: Rank-1 accuracy mAP
Baselines: 3DSL GI-ReID CAL AIM LDF 3DInvarReID SCNet CVSL CLIP3DReID CCPG TransReID Instruct-ReID
Builds on: CLIP EVA-CLIP gradient reversal layer CAL CLIP3DReID
Closest papers by meaning
- Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification (CVPR 2025 · similarity 0.91)
- 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)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.85)
- Person De-reidentification: A Variation-guided Identity Shift Modeling (CVPR 2025 · similarity 0.85)
- Distilling CLIP with Dual Guidance for Learning Discriminative Human Body Shape Representation (CVPR 2024 · 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.