Domain-adaptive Person Re-identification without Cross-camera Paired Samples [arxiv]

Framework of our method

Usage

• This project is based on TransReID[1] (paper and official code)
• Usage of this code is free for research purposes only.

Requirements

• python = 3.9
• torch = 1.7.1
• torchvision = 0.8.2
• opencv-python = 4.5.1.48
• timm = 0.4.5
• yacs = 0.1.8

We use a single NVIDIA RTX3090 GPU for training and evaluation.

Prepare Datasets

mkdir data

Download the person datasets Market1501, MSMT17, CUHK03.

you can find the Market1501-SCT* in [SCT-ReID] [2], and MSMT17-SCT* in [CCFP] [3].

Make new directories in data and organize them as follows:

data
├── market1501
│       └── bounding_box_train
│       └── bounding_box_train
│       └── bounding_box_train_sct
│       └── bounding_box_test
│       └── query

├── CUHK03
│       └── labeled
│            └── bounding_box_train
│            └── bounding_box_test
│            └── query
│       └── detected
│            └── bounding_box_train
│            └── bounding_box_test
│            └── query

├── MSMT17
│       └── test
│       └── train
│       └── list_gallery.txt
│       └── list_query.txt
│       └── list_val.txt

Tips:

Market:

the file "bounding_box_train_sct" is the training set for market-SCT

When use different protocol dataset for train, please change the train_dir in dataset files:

Market:

market1501.py

    self.train_dir = osp.join(self.dataset_dir, 'bounding_box_train')
  # self.train_dir = osp.join(self.dataset_dir, 'bounding_box_train_sct')

MSMT17:

msmt17.py

   self.list_train_path = osp.join(self.dataset_dir, 'list_train.txt')
 # self.list_val_path = osp.join(self.dataset_dir, 'list_train_sct.txt')

Training

• 1.Replace dataset path with your own path in vit_3_domain.yml

DATASETS:
  NAMES: ('') # source data
  TARGET: ('') # target data
  ROOT_DIR: ('')

• 2.Begin the training

  python train.py

Test

• 1.Replace test file path with your own path in vit_3_domain.yml

TEST:
  WEIGHT: '*.pth'

• 2.Replace dataset path with your own path in vit_3_domain.yml

# keep the source & target data are same as the *.pth
DATASETS:
  NAMES: ('') # source data
  TARGET: ('') # target data
  ROOT_DIR: ('')

• 3.Change the num_class in test.py

# Market source num_class = 751
model, _, _，_,_ = make_model(cfg, num_class = 751, num_class_t = c_pids, camera_num = camera_num, view_num = view_num)

• 4.Begin the test

  python test.py

Contact

If you have any question, please feel free to contact us. E-mail: maoyanmeicy@163.com

Prepare ViT Pre-trained Models

You need to download the ImageNet pretrained transformer model : ViT-Base, ViT-Small, DeiT-Small, DeiT-Base

Reference

[1]S. He, H. Luo, P. Wang, F. Wang, H. Li, and W. Jiang, “Transreid: Transformer-based object re-identification,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp.15013–15022.
[2]T. Zhang, L. Xie, L. Wei, Y. Zhang, B. Li, and Q. Tian, “Single camera training for person re-identification,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol.34, no.07, 2020, pp.12878–12885.
[3]W. Ge, C. Pan, A. Wu, H. Zheng, and W.-S. Zheng, “Cross-camera feature prediction for intra-camera supervised person re-identification across distant scenes,” in Proceedings of the 29th ACM International Conference on Multimedia, 2021, pp. 3644–3653.

Citation

If you find this code useful for your research, please cite our paper

@ARTICLE{UDA2023,
  title={Domain-adaptive Person Re-identification without Cross-camera Paired Samples},
  author={Huafeng Li, Yanmei Mao, Yafei Zhang, Guanqiu Qi, and Zhengtao Yu},
  journal={arXiv preprint arXiv:2307.06533},
  year={2023}
}