Official implementation of Paper ''A Versatile Framework for Multi-scene Person Re-identification''.
2024/3/19: Our arxiv paper can be found here
2024/3/16: Code and Pre-trained model are released. Check Baidu Drive for the pre-trained model (key: pami)
Please visit the following link to download the dataset.
- Market-1501
- MSMT17
- MLR-CUHK03, key: pami
- Cele-ReID
- PRCC
- Occ-Duke
- SYSU-mm01
Once you download the datasets, make sure to modify the dataset's root manually here and here.
Please follow TransReID to configure the running environment.
We provide our used package list in full-environment.txt for a reference.
-
Download the pre-trained model at Baidu Drive, the key is pami.
-
Create a directory named ckpts and then put the downloaded model into it, or you can modify the MODEL.PRETRAIN_PATH in
./bash/run_VersReID.sh
to your own pre-trained model path. -
To reproduce the results in our paper, just simply run this script:
bash ./bash/run_VersReID.sh
You can modify the configs by yourself to explore more settings.
Note: Training the ReID-Bank requires ~20G GPU Memory, V-Branch requires ~30G GPU Memory. We highly recommend use cuda 10.2 for better reproducibility. Moreover, the code does not support multi-GPU training currently.
If you have any problem, feel free to open an issue or contact me :-)
-
This repository is heavily based on TransReID, many thanks to the authors.
-
If you find this repo helpful, please consider citing us:
@article{zheng2024versreid,
title = {A Versatile Framework for Multi-scene Person Re-identification},
author = {Zheng, Wei-Shi and Yan, Junkai and Peng, Yi-Xing},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
year = {2024}
}