Introduction

This is the implementation codes for Exploring Visual Context for Weakly Supervised Person Search

Overall pipeline of the proposed context-guided feature learning framework for weakly supervised person search. We aim to build a framework for person search with only bounding box annotations, where the re-id embeddings are jointly learned with detection. Without identity annotations, initial pseudo labels (colored points) are generated with ImageNet-pretrained weights. We employ the detection context to pull features belonging to the same identity together, while pushing the re-id features of the pedestrians away from the background features. A hard-negative mining strategy is designed to effectively employ the information in the memory. We use the scene context to generate more accurate clustering results.

Installation

The project is based on MMdetection, please refer to install.md to install MMdetection.

We utilized mmcv=1.2.6, pytorch=1.7.1

Dataset

Download CUHK-SYSU and PRW.

We provide coco-style annotation in demo/anno.

For CUHK-SYSU, change the path of your dataset and the annotaion file in the config file L2, L35, L40, L46, L51

For PRW, change the path of your dataset and the annotaion file in the config file L2, L35, L40, L46, L51

Experiments

1. Train

cd jobs/cuhk/
sh train.sh

2. Test CUHK-SYSU Download trained CUHK checkpoint.

cd jobs/cuhk/
sh test.sh

3. Train PRW

cd jobs/prw/
sh train.sh

4. Test PRW Download trained PRW checkpoint. Change the paths in L125 in test_results_prw.py

cd jobs/prw
sh test.sh

Performance

Dataset	Model	mAP	Rank1	Config	Link
CUHK-SYSU	CGPS	80.1%	82.1%	cfg	model
PRW	CGPS	16.6%	68.2%	cfg	model

Reference Codes

Thanks for the great projects of MMdetection, OpenUnReID and AlignPS.

License

This project is released under the Apache 2.0 license.

Citation

If you use this project in your research, please cite this project.

@misc{yan2021exploring,
      title={Exploring Visual Context for Weakly Supervised Person Search}, 
      author={Yichao Yan and Jinpeng Li and Shengcai Liao and Jie Qin and Bingbing Ni and Xiaokang Yang and Ling Shao},
      year={2021},
      eprint={2106.10506},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}