shijing123 / paul

This repo contains the source code for our CVPR'19 work "Patch-based discriminative feature learning for unsupervised person re-identification"

Our implementation is based on Pytorch.

Prerequisites

1. Pytorch 0.4.1
2. Python 3.6

Preparation

• Data preparation

mkdir data

ln -s [PATH TO MSMT17_V1] ./data/MSMT17_V1
ln -s [PATH TO DUKE] ./data/DukeMTMC-reID
ln -s [PATH TO Market] ./data/Market

• set the path of ImageNet pretrained models

ln -s [THE PATH OF IMAGENET PRE-TRAINED MODELS] imagenet_models

Run the code

• For pretraining the model

cd ./src/train
supervised_train.py --gpu [CHOOSE WHICH GPU TO RUN] --exp-name [YOUR EXP NAME]

• Optionally, you can download our pretrained model from google drive or baidu cloud (kvkz) and place it in ./snapshot/MSMT17_PRE/

mkdir ./snapshot
mkdir ./snapshot/MSMT17_PRE
cp [PATH TO PRETRAINED MODEL] ./snapshot/MSMT17_PRE/
# it means the name of the experiment of pretraining is 'MSMT17_PRE'  

• For unsupervised training

cd ./src/unsupervised

# for market
python unsupervised_train.py --data MARKET --gpu [CHOOSE WHICH GPU TO RUN] \
--pre-name [THE EXP NAME OF PRE-TRIANED MODEL] --exp-name [YOUR EXP NAME] \
--batch-size 42 --scale 15 --lr 0.0001 

 # for duke
python unsupervised_train.py --data DUKE --gpu [CHOOSE WHICH GPU TO RUN] \
--pre-name [THE EXP NAME OF PRE-TRIANED MODEL] --exp-name [YOUR EXP NAME] \
--batch-size 40 --scale 5 --lr 0.0001 

Reference

If you find our work helpful in your research, please kindly cite our paper:

Qize Yang, Hong-Xing Yu, Ancong Wu, Wei-Shi Zheng, "Patch-based discriminative feature learning for unsupervised person re-identification", In CVPR, 2019.

bib:

@inproceedings{yu2019unsupervised,
  title={Patch-based discriminative feature learning for unsupervised person re-identification},
  author={Qize, Yang and Yu, Hong-Xing and Wu, Ancong and Zheng, Wei-Shi},
  year={2019},
  booktitle={IEEE International Conference on Computer Vision and Pattern Recognition (CVPR)},
}

If you have any problem/question, please feel free to contact me at [email protected] or open an issue. Thank you.