Research map / Label-efficient and robust learning / Recognition and long-tail

Exploring the Camera Bias of Person Re-identification

ICLR 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Cross-modality and occlusion re-identification · 0 citations (OpenAlex)

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.

Shows person ReID models carry strong camera bias on unseen domains and in unsupervised training, revisits camera-specific feature normalization as a test-time debiasing fix, and proposes two USL training strategies (debiased pseudo-labeling and discarding single-camera clusters) that boost existing methods.

Problem

Person re-identification models are biased toward camera labels, a problem studied mainly on training domains; camera bias on unseen domains and in unsupervised learning remains under-explored.

Approach

Revisit simple camera-specific feature normalization (per-camera mean centering and std scaling) at test time, analyze why it works (camera-sensitive dimensions, low-level properties, body angle), and use the same normalization for debiased pseudo-labeling plus discarding single-camera clusters in unsupervised training.

Key results

• Camera-specific normalization on TransReID-SSL trained on MSMT17 raises Market-1501 mAP from 53.6 to 62.3 and R1 from 78.1 to 83.5.
• Applying the proposed USL strategies to CC raises MSMT17 mAP from 29.8 to 49.1 (about 19.3 points).
• PPLR-CAM* + Ours reaches 44.4 mAP on MSMT17 versus 40.7 for the reproduced PPLR-CAM baseline.

Measured effects reported in the paper

Surprising findings

• Camera-specific normalization gives no improvement, and can slightly hurt, on the seen domain for camera-aware or supervised models.
• Camera bias persists even when training data has balanced cameras (CC trained on PersonX with ground-truth labels still gains from normalization).

Limitations

• Normalization requires per-camera mean and variance statistics, which scale with data and can be memory-heavy for many samples per camera.
• Camera-aware and supervised models gain nothing on seen domains, and the debiasing gains depend on the test domain.
• Discarding single-camera clusters removes a large share of samples early in training (63.5% at epoch 0).

Experimental setup

Datasets: MSMT17 Market-1501 CUHK03-NP PersonX VeRi-776

Benchmarks: Market-1501 MSMT17 CUHK03-NP PersonX VeRi-776

Metrics: mAP Rank-1 Rank-5 Rank-10 NMI bias NMI accuracy

Baselines: CC PPLR TransReID-SSL ISR PPLR-CAM TransReID SOLIDER CAP ICE-CAM CAJ PAT SPCL

Builds on: PPLR CC SPCL TransReID DBSCAN

Closest papers by meaning

• Adversarially Occluded Samples for Person Re-Identification (CVPR 2018 · similarity 0.91)
• Person30K: A Dual-Meta Generalization Network for Person Re-Identification (CVPR 2021 · similarity 0.90)
• Single Camera Training for Person Re-Identification (AAAI 2020 · similarity 0.90)
• Generalizable Person Re-Identification by Domain-Invariant Mapping Network (CVPR 2019 · similarity 0.90)
• From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.89)
• Person De-reidentification: A Variation-guided Identity Shift Modeling (CVPR 2025 · similarity 0.89)