Research map / Label-efficient and robust learning / Recognition and long-tail
Rethinking the Distribution Gap of Person Re-identification with Camera-based Batch Normalization
Zijie Zhuang, Longhui Wei, Lingxi Xie, Tianyu Zhang, Hengheng Zhang, Haozhe Wu, Haizhou Ai, Qi Tian
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace dataset-based formulation with camera-based batch normalization (CBN) | ReID ResNet-50 baseline · Market to Duke direct transfer | Rank-1: 37.0 -> 58.7 | Improved |
| replace dataset-based formulation with CBN | ReID ResNet-50 baseline · Duke to Market direct transfer | Rank-1: 53.2 -> 72.7 | Improved |
| replace dataset-based formulation with CBN | ReID ResNet-50 baseline · Duke fully-supervised | Rank-1: 81.5 -> 82.5 | Improved |
| integrate CBN into BoT* (Random Erasing disabled) | BoT* · Market fully-supervised | Rank-1: 93.6 -> 94.3 | Improved |
| integrate CBN into ECN domain adaptation | ECN · Duke to Market | Rank-1: 75.1 -> 81.7 | Improved |
| add inter-camera annotations (fully versus weakly supervised) under CBN | CBN ReID · Market train, Market and MSMT test | Rank-1: 83.3 -> 91.3 (Market); 26.8 -> 25.3 (MSMT) | Mixed |
| use CBN instead of conventional formulation for Data-Free incremental learning (length 2) | baseline ReID · Market first dataset retention | Rank-1: 82.2% -> 88.3% | Improved |
| use CBN instead of conventional formulation with Replay (length 3) | baseline ReID · MSMT first dataset retention | mAP: 60.9% -> 76.0% | Improved |
| add CBN at training time on top of CBN at test time (ablation) | BN-trained model with CBN test · Duke to Market | Rank-1: 69.5 -> 72.7 | Improved |
| replace AdaBN test-set alignment with CBN camera-wise alignment at test (ablation) | BN-trained model · Duke to Market | Rank-1: 55.8 -> 69.5 | Improved |
| replace BN with IBN (ablation) | ReID baseline · Duke to Duke and Duke to Market | Rank-1: 81.5 -> 77.6 (Duke); 53.2 -> 61.7 (Market) | Mixed |
| estimate camera statistics from 50 batches instead of 1 (ablation) | CBN · Market to Duke | mAP: 37.34 -> 38.21 | Improved |
Closest papers by meaning
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.92)
- Single Camera Training for Person Re-Identification (AAAI 2020 · similarity 0.90)
- A Versatile Framework for Multi-Scene Person Re-Identification (TPAMI 2024 · similarity 0.90)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.88)
- CLIMB-ReID: A Hybrid CLIP-Mamba Framework for Person Re-Identification (AAAI 2025 · similarity 0.88)
- Person De-reidentification: A Variation-guided Identity Shift Modeling (CVPR 2025 · similarity 0.88)
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.
