Skip to content

Latest commit

 

History

History
94 lines (73 loc) · 2.78 KB

README.md

File metadata and controls

94 lines (73 loc) · 2.78 KB

Python >=3.5 PyTorch >=1.0

Group Sampling

Rethinking Sampling Strategies for Unsupervised Person Re-identification
Xumeng Han, Xuehui Yu, Guorong Li, Jian Zhao, Gang Pan, Qixiang Ye, Jianbin Jiao and Zhenjun Han
IEEE Transactions on Image Processing (TIP) 2023 (arXiv:2107.03024)

Requirements

Installation

git clone https://github.com/wavinflaghxm/GroupSampling.git
cd GroupSampling
python setup.py develop

Prepare Datasets

cd examples && mkdir data

Download the person datasets Market-1501, DukeMTMC-reID, MSMT17. Then unzip them under the directory like:

GroupSampling/examples/data
├── market1501
│   └── Market-1501-v15.09.15
├── dukemtmc
│   └── DukeMTMC-reID
└── msmt17
    └── MSMT17_V2

Training

We utilize 1 GTX-2080TI GPU for training.

  • Use --group-n 256 for Market-1501, --group-n 128 for DukeMTMC-reID, and --group-n 1024 for MSMT17.

Market-1501:

CUDA_VISIBLE_DEVICES=0 python examples/train.py -d market1501 --logs-dir logs/market_resnet50 --group-n 256

DukeMTMC-reID:

CUDA_VISIBLE_DEVICES=0 python examples/train.py -d dukemtmc --logs-dir logs/duke_resnet50 --group-n 128

MSMT17:

CUDA_VISIBLE_DEVICES=0 python examples/train.py -d msmt17 --logs-dir logs/msmt_resnet50 --group-n 1024 --iters 800

We recommend using 4 GPUs to train MSMT17 for better performance.

CUDA_VISIBLE_DEVICES=0,1,2,3 python examples/train.py -d msmt17 --logs-dir logs/msmt_resnet50-gpu4 --group-n 1024 -b 256 --momentum 0.1 --lr 0.00005

Evaluation

To evaluate the model, run:

CUDA_VISIBLE_DEVICES=0 python examples/test.py -d $DATASET --resume $PATH

Some examples:

### Market-1501 ###
CUDA_VISIBLE_DEVICES=0 python examples/test.py -d market1501 --resume logs/market_resnet50/model_best.pth.tar

Results

results

Citation

If you find this work useful for your research, please cite:

@article{han2022rethinking,
  title={Rethinking Sampling Strategies for Unsupervised Person Re-Identification}, 
  author={Han, Xumeng and Yu, Xuehui and Li, Guorong and Zhao, Jian and Pan, Gang and Ye, Qixiang and Jiao, Jianbin and Han, Zhenjun},
  journal={IEEE Transactions on Image Processing}, 
  year={2023},
  volume={32},
  pages={29-42},
  doi={10.1109/TIP.2022.3224325}}

Acknowledgements

Codes are built upon SpCL. Thanks to Yixiao Ge for opening source.