Self-Supervised Pre-Training for Transformer-Based Person Re-Identification [pdf]

The official repository for Self-Supervised Pre-Training for Transformer-Based Person Re-Identification.

Requirements

Installation

pip install -r requirements.txt

We recommend to use /torch=1.7.1 /torchvision=0.8.2 /timm=0.3.4 /cuda>10.1 /faiss-gpu=1.7.1/ 16G or 32G V100 for training and evaluation. If you find some packages are missing, please install them manually. You can refer to DINO, TransReID and cluster-contrast-reid to install the environment of pre-training, supervised ReID and unsupervised ReID, respectively.

Prepare Datasets

mkdir data

Download the datasets:

• Market-1501
• MSMT17
• LUPerson. We don't have the copyright of the LUPerson dataset. Please contact authors of LUPerson to get this dataset.
• You can download the file list ordered by the CFS score for the LUPerson. [CFS_list.pkl]

Then unzip them and rename them under the directory like

data
├── market1501
│   └── bounding_box_train
│   └── bounding_box_test
│   └── ..
├── MSMT17
│   └── train
│   └── test
│   └── ..
└── LUP
    └── images 
    └── CFS_list.pkl 

Pre-trained Models

Model	Download
ViT-S/16	link
ViT-S/16+ICS	link
ViT-B/16+ICS	link

Please download pre-trained models and put them into your custom file path.

ReID performance

We have reproduced the performance to verify the reproducibility. The reproduced results may have a gap of about 0.1~0.2% with the numbers in the paper.

Supervised ReID

Market-1501

Model	Image Size	Paper	Reproduce	Download
ViT-S/16	256*128	91.0/96.0	91.2/95.8	model / log
ViT-S/16+ICS	256*128	91.3/96.2	91.4/96.2	model / log
ViT-B/16+ICS	384*128	93.2/96.7	93.1/96.6	model / log

MSMT17

Model	Image Size	Paper	Reproduce	Download
ViT-S/16	256*128	66.1/84.6	66.3/84.8	model / log
ViT-S/16+ICS	256*128	68.1/86.1	68.3/86.2	model / log
ViT-B/16+ICS	384*128	75.0/89.6	75.1/89.6	model / log

USL ReID

Market-1501

Model	Image Size	Paper	Reproduce	Download
ViT-S/16	256*128	88.2/94.2	88.4/94.6	model / log
ViT-S/16+ICS	256*128	89.6/95.3	89.5/95.3	model / log

MSMT17

Model	Image Size	Paper	Reproduce	Download
ViT-S/16	256*128	40.9/66.4	40.9/66.4	model / log
ViT-S/16+ICS	256*128	50.6/75.0	50.6/75.0	model / log

UDA ReID

MSMT2Market

Model	Image Size	Paper	Reproduce	Download
ViT-S/16	256*128	89.4/95.4	89.2/95.3	model / log
ViT-S/16+ICS	256*128	89.9/95.5	89.9/95.4	model / log

Market2MSMT

Model	Image Size	Paper	Reproduce	Download
ViT-S/16	256*128	47.4/70.8	47.7/71.2	model / log
ViT-S/16+ICS	256*128	57.8/79.5	57.8/79.4	model / log

Acknowledgment

Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.

LUPerson, DINO, TransReID, cluster-contrast-reid.

Citation

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

@article{luo2021self,
  title={Self-Supervised Pre-Training for Transformer-Based Person Re-Identification},
  author={Luo, Hao and Wang, Pichao and Xu, Yi and Ding, Feng and Zhou, Yanxin and Wang, Fan and Li, Hao and Jin, Rong},
  journal={arXiv preprint arXiv:2111.12084},
  year={2021}
}

Contact

If you have any question, please feel free to contact us. E-mail: michuan.lh@alibaba-inc.com or haoluocsc@zju.edu.cn