Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification CVPR 2019

Preparation

Requirements: Python=3.6 and Pytorch>=1.0.0

1. Install Pytorch

2. Download dataset

   • Market-1501 [BaiduYun] [GoogleDriver] CamStyle (generated by CycleGAN) [GoogleDriver] [BaiduYun] (password: 6bu4)

   • DukeMTMC-reID [BaiduYun] (password: bhbh) [GoogleDriver] CamStyle (generated by CycleGAN) [GoogleDriver] [BaiduYun] (password: 6bu4)

   • MSMT17 + CamStyle (generated by StarGAN) [BaiduYun] (password: 6bu4) [GoogleDriver] We reformulate the structure of MSMT17 the same as Market-1501.

   • Unzip each dataset and corresponding CamStyle under ‘ECN/data/‘

   Ensure the File structure is as follow:

   ECN/data    
   │
   └───market OR duke OR msmt17
      │   
      └───bounding_box_train
      │   
      └───bounding_box_test
      │   
      └───bounding_box_train_camstyle
      | 
      └───query

Training and test domain adaptation model for person re-ID

  # For Duke to Market-1501
  python main.py -s duke -t market --logs-dir logs/duke2market-ECN
  # For Market-1501 to Duke
  python main.py -s market -t duke --logs-dir logs/market2duke-ECN
  # For Market-1501 to MSMT17
  python main.py -s market -t msmt17 --logs-dir logs/market2msmt17-ECN --re 0
  # For Duke to MSMT17
  python main.py -s duke -t msmt17 --logs-dir logs/duke2msmt17-ECN --re 0

Results

References

• [1] Our code is conducted based on open-reid

• [2] Camera Style Adaptation for Person Re-identification. CVPR 2018.

• [3] Generalizing A Person Retrieval Model Hetero- and Homogeneously. ECCV 2018.

Citation

If you find this code useful in your research, please consider citing:

@inproceedings{zhong2019invariance,
  title={Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification},
  author={Zhong, Zhun and Zheng, Liang and Luo, Zhiming and Li, Shaozi and Yang, Yi},
  booktitle={Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2019},
}

Contact me

If you have any questions about this code, please do not hesitate to contact me.

Zhun Zhong