Research map / Label-efficient and robust learning / Recognition and long-tail
CDE-Learning: Camera Deviation Elimination Learning for Unsupervised 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 CDE-Learning, an unsupervised person Re-ID framework that aligns camera-domain centroids before DBSCAN clustering, uses a camera-PK multi-prototype contrastive loss, and adds a same-camera prediction auxiliary task.
Problem
Unsupervised person re-identification suffers from camera-induced feature deviations (contrast, saturation, resolution) that cause false positives in pseudo-label assignment and require hand-tuned camera constraint hyperparameters.
Approach
Refines features by shifting each camera domain toward the global centroid before clustering (CDC), unifies intra- and inter-camera contrastive constraints into one loss built from a multi-prototype dictionary without extra hyperparameter tuning (CDB), and adds a binary same-camera prediction auxiliary task (CAA).
Key results
- Achieves 87.4% mAP on Market-1501, versus 86.5% for the best purely unsupervised baseline (DCMIP).
- Achieves 43.4% mAP on MSMT17 and 86.2% mAP on PersonX.
- Ablation on Market-1501: CDC alone raises mAP from 82.8% to 86.1%; full combination reaches 87.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CDC to the baseline (ablation) | CDE-Learning · Market-1501 | mAP: 82.8 to 86.1 | Helped |
| Add CDC to the baseline (ablation) | CDE-Learning · MSMT17 | mAP: 33.0 to 40.4 | Helped |
| Add CDB to the baseline (ablation) | CDE-Learning · Market-1501 | mAP: 82.8 to 85.4 | Helped |
| Add CAA to the baseline (ablation) | CDE-Learning · MSMT17 | mAP: 33.0 to 36.5 | Helped |
| Use CDC, CDB, and CAA together (ablation) | CDE-Learning · Market-1501 | mAP: 82.8 to 87.4 | Helped |
| Use CDC, CDB, and CAA together (ablation) | CDE-Learning · MSMT17 | mAP: 33.0 to 43.4 | Helped |
| Replace deviation balancing loss with inter-camera loss (ablation) | CDE-Learning · Market-1501 | mAP: 87.4 -> 85.3 | Hurt |
| Replace deviation balancing loss with intra-camera loss (ablation) | CDE-Learning · Market-1501 | mAP: 87.4 -> 86.1 | Hurt |
| Fine-tune separate intra- and inter-camera losses instead of using deviation balancing loss (ablation) | CDE-Learning · Market-1501 | mAP: 87.4 -> 86.9 | Hurt |
| Concatenate random feature pairs instead of same-pseudo-label pairs (ablation) | CAA · Market-1501 | mAP: 87.4 -> 86.8 | Hurt |
| Vary temperature τ (ablation) | CDE-Learning · Market-1501 | mAP: 83.9, 86.5, 87.4, 86.1, 84.8, 82.3 | Mixed |
Surprising findings
- On MSMT17, the method's Rank-1 (70.5%) is below several camera-aware baselines such as PPLR (71.8%) and NPSS (72.1%), despite higher mAP.
Limitations
- Relies on camera labels being available, which the method uses for both alignment and constraints.
- Temperature and PK sampling parameters are sensitive; performance drops noticeably outside a narrow range (e.g., tau=0.3 gives 82.3% mAP).
Experimental setup
Datasets: Market-1501 MSMT17 PersonX CUHK03
Benchmarks: Market-1501 MSMT17 PersonX CUHK03
Metrics: mAP Rank-1 Rank-5 Rank-10
Baselines: JVTC IICS CAP ICE PPLR CDR NPSS HCACE CGC MCRN SECRET CCL
Builds on: DBSCAN MoCo PK sampling ICE PPLR
Closest papers by meaning
- Joint Noise-Tolerant Learning and Meta Camera Shift Adaptation for Unsupervised Person Re-Identification (CVPR 2021 · similarity 0.89)
- Intra-Inter Camera Similarity for Unsupervised Person Re-Identification (CVPR 2021 · similarity 0.88)
- Camera-Aware Proxies for Unsupervised Person Re-Identification (AAAI 2021 · similarity 0.88)
- Asymmetric Co-Teaching for Unsupervised Cross-Domain Person Re-Identification (AAAI 2020 · similarity 0.87)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.87)
- Person De-reidentification: A Variation-guided Identity Shift Modeling (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.