This is the official repo for the CVPR 2021 paper "MetaCam+DSCE".

[2021.5.24] We recorded a video on Zhidongxi.

• CUDA>=10.0

• At least two 1080-Ti GPUs

• Other necessary packages listed in requirements.txt

• Training Data

  (Market-1501, DukeMTMC-reID and MSMT-17. You can download these datasets from Zhong's repo)

  Unzip all datasets and ensure the file structure is as follow:

  MetaCam_DSCE/data    
  │
  └───market1501 OR dukemtmc OR msmt17
       │   
       └───DukeMTMC-reID OR Market-1501-v15.09.15 OR MSMT17_V1
           │   
           └───bounding_box_train
           │   
           └───bounding_box_test
           | 
           └───query
           │   
           └───list_train.txt (only for MSMT-17)
           | 
           └───list_query.txt (only for MSMT-17)
           | 
           └───list_gallery.txt (only for MSMT-17)
           | 
           └───list_val.txt (only for MSMT-17)
  

See run.sh for details.

This repo borrows partially from MWNet (meta-learning), ECN (exemplar memory) and SpCL (faiss-based acceleration). If you find our code useful, please cite their papers.

1. Pre-trained MMT-500 models to reproduce Tab. 3 of our paper. BaiduNetDisk, Passwd: jr1l. Google Drive.

2. Pedestrian images used to plot Fig.3 in our paper. BaiduNetDisk, Passwd: f248. Google Drive.

   Please download 'marCam' and 'dukeCam', put them under 'MetaCam_DSCE/data', uncomment L#87-89 and L#163-168 of train_usl_knn_merge.py to visualize pedestrian features.

3. Training logs. BaiduNetDisk, Passwd: mecq. Google Drive.

@inproceedings{yang2021joint,
  title={Joint Noise-Tolerant Learning and Meta Camera Shift Adaptation for Unsupervised Person Re-Identification},
  author={Yang Fengxiang and Zhong Zhun and Luo Zhiming and Cai Yuanzheng and Lin Yaojin and Li Shaozi and Nicu Sebe},
  booktitle={CVPR},
  pages={4855--4864},
  year={2021}
}

Email: yangfx@stu.xmu.edu.cn