Context-Aware Unsupervised Clustering for Person Search

Official pytorch implementation of Context-Aware Unsupervised Clustering for Person Search at BMVC 2021.

Settings

1. Clone the repository

   git clone https://github.com/VIP-Lab-UNIST/CUCPS.git && cd CUCPS

2. Build environment with conda

   conda env create --prefix <your_conda_env_path> -f environment.yml

3. Set the Datasets

   Download CUHK-SYSU and PRW to your location and set the dataset path in ./lib/datasets/__init__.py.

Train

Set the save path(--path) in the .sh files.

1. Train CUHK-SYSU

   ./train-cuhk.sh

2. Train PRW

   ./train-prw.sh

Test

Add the checkpoint that you want to evaluate on the --checkpoint_name option in test-*.sh files. The model shold be with args.json file.

• You can also evaluate pretrained weight of test 26, 18 epochs for CUHK-SYSU, PRW, respectively.

1. Test CUHK-SYSU

   ./test-cuhk.sh

2. Test PRW(regular gallery)

   ./test-prw.sh

3. Test PRW(multi-view gallery)

   Set ignore_cam_id=False and remove_unlabel=False of search_performance_calc function on runs/test.py

   ./test-prw.sh

The result will be saved in same directory with checkpoint file.

Results on the pretrained models (paper results)

Datasets	mAP(%)	Rank-1(%)
CUHK-SYSU	81.1	83.2
PRW(regular gallery)	41.7	86.0
PRW(multi-view gallery)	36.6	64.9

Citation

@article{cucps,
  title={Context-Aware Unsupervised Clustering for Person Search},
  author={Byeong-Ju Han*, Kuhyeun Ko* and Jae-Young Sim},
  booktitle = {British Machine Vision Conference (BMVC)},
  year={2021},
}