Research map / Label-efficient and robust learning / Recognition and long-tail
Divide and Conquer: Hybrid Pre-training for Person Search
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 a hybrid pre-training framework that uses detection and re-ID sub-task data (with or without labels) to pre-train person search models, improving over ImageNet initialization.
Problem
Person search methods rely on ImageNet pre-trained backbones, which mismatch the detection plus re-ID nature of the task, and no large-scale person search pre-training dataset exists.
Approach
Pre-train a full person search model using only sub-task data by combining fully supervised detection and re-ID losses (with detection images given pseudo identities) and SimSiam-style self-supervised contrastive learning on unlabeled images, plus an intra-task adversarial alignment module (IAM) that aligns features across datasets of the same task.
Key results
- On PRW with ResNet50 NAE, pre-trained model raises mAP from 42.67 to 47.08 (+10.3% relative).
- ROI-AlignPS with the pre-trained model reaches 95.4 mAP on CUHK-SYSU and 54.5 mAP on PRW, a 2.7pp gain over the re-implemented baseline on PRW.
- Pre-trained NAE with only 60% of PRW target data outperforms the ImageNet-initialized NAE trained on all annotations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace ImageNet pre-training with the proposed pre-training | NAE, AlignPS, and ROI-AlignPS · PRW | mAP: 42.67→47.08 (↑4.41); 45.90→49.14 (↑3.24); 51.84→54.47 (↑2.63) | Helped |
| Replace ImageNet pre-training with the proposed pre-training | ROI-AlignPS · CUHK-SYSU and PRW | mAP: 95.33→95.38; 51.84→54.47 | Helped |
| Use the proposed pre-training instead of ImageNet pre-training | NAE with ResNet50, ResNet18, and ShuffleNet · PRW | mAP: 42.67→47.08 (↑10.3%); 35.87→39.79 (↑10.9%); 34.29→40.98 (↑19.5%) | Helped |
| Use expand_resize instead of random_paste for re-ID data unification (ablation) | NAE · PRW | mAP: 46.49; 47.08 | Helped |
| Use expand_resize instead of random_paste for re-ID data unification (ablation) | NAE · PRW | Top-1: 83.36; 84.00 | Helped |
| Add inter-task alignment to intra-task alignment (ablation) | Proposed pre-training method · PRW | mAP: 47.08; 46.53 | Hurt |
| Add inter-task alignment to intra-task alignment (ablation) | Proposed pre-training method · PRW | Top-1: 84.00; 83.16 | Hurt |
| Add the domain-classifier consistency regularizer (ablation) | Intra-task alignment module · PRW | mAP: 0.3pp | Helped |
| Use the proposed pre-trained model instead of ImageNet initialization | ROI-AlignPS · PoseTrack21 | mAP: 59.21→63.62 | Helped |
| Use the proposed pre-trained model instead of ImageNet initialization | ROI-AlignPS · PoseTrack21 | Top-1: 82.10→87.13 | Helped |
| Use the proposed pre-trained model instead of ImageNet initialization | ROI-AlignPS · CUHK-SYSU | mAP: 95.33→95.38 | Helped |
| Use the proposed pre-trained model instead of ImageNet initialization | ROI-AlignPS · PRW | mAP: 51.84→54.47 | Helped |
Surprising findings
- Inter-task alignment between detection and re-ID features slightly hurts performance (47.08 to 46.53 mAP), so only intra-task alignment is used.
- Gains on CUHK-SYSU are negligible (+0.05 mAP) because that benchmark is near saturation.
Limitations
- Gains shrink as more pre-training datasets are added (2.12pp at 4 datasets to 0.6pp at 5 datasets).
- Pre-training uses only about 106k images, roughly 20x PRW, which is limited relative to large-scale pre-training.
- Gain on CUHK-SYSU is marginal because the benchmark is saturated.
Experimental setup
Datasets: CrowdHuman EuroCity Persons MSMT17 CUHK03 LUPerson CUHK-SYSU PRW PoseTrack21
Benchmarks: CUHK-SYSU PRW PoseTrack21
Metrics: mAP Top-1
Baselines: ImageNet pre-trained ResNet50 NAE AlignPS ROI-AlignPS SeqNet PSTR COAT OR TCTS
Builds on: SimSiam NAE OIM AlignPS ROI-AlignPS
Closest papers by meaning
- Diverse Knowledge Distillation for End-to-End Person Search (AAAI 2021 · similarity 0.90)
- Mixture of Submodules for Domain Adaptive Person Search (CVPR 2025 · similarity 0.86)
- Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification (CVPR 2025 · similarity 0.86)
- Multi-Prompts Learning with Cross-Modal Alignment for Attribute-Based Person Re-identification (AAAI 2024 · similarity 0.86)
- Diverse Person: Customize Your Own Dataset for Text-Based Person Search (AAAI 2024 · similarity 0.85)
- Domain Generalizable Person Search Using Unreal Dataset (AAAI 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.