FlyingRoastDuck / ACT_AAAI20

code for AAAI 2020 paper "ACT"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Asymmetric Co-Teaching for Unsupervised Cross Domain Person Re-Identification (AAAI 2020)

Code for AAAI 2020 paper Asymmetric Co-Teaching for Unsupervised Cross Domain Person Re-Identification.

Results

Requirements

  • python 3.7
  • Server with 4 GPUs
  • Market1501, DukeMTMC-reID and other datasets.
  • Other necessary packages listed in requirements.txt

Adaptation with ACT

  1. Download all necessary datasets and move them to 'data' by following instructions in 'data/readme.md'

  2. If you want to train from the pre-adapted model for fast reproduction, please download all models in Resources and run the following command:

python selftrainingACT.py  --src_dataset {src_dataset_name} --tgt_dataset {tgt_dataset_name} --resume {model's path} --data_dir ./data --logs_dir {path to save model}

avaliable choices to fill "src_dataset_name" and "tgt_dataset_name" are: market1501 (for Market1501), dukemtmc (for DukeMTMC-reID), cuhk03 (for CUHK03).

  1. If you want to train from scratch, please train source model and adapted model by using code in Adaptive-ReID and follow #2.

Adaptation with other co-teaching-like structures.

To reproduce the results in Tab. 2 of our paper, please run selftrainingRCT.py and selftrainingCT.py in similar way.

Adaptation with other clustering methods.

To reproduce Tab.3 of our paper, run selftrainingKmeans.py (co-teaching version) and selftrainingKmeansAsy.py (ACT version).

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

@inproceedings{yang2020asymmetric,
  title={Asymmetric Co-Teaching for Unsupervised Cross-Domain Person Re-Identification.},
  author={Yang, Fengxiang and Li, Ke and Zhong, Zhun and Luo, Zhiming and Sun, Xing and Cheng, Hao and Guo, Xiaowei and Huang, Feiyue and Ji, Rongrong and Li, Shaozi},
  booktitle={AAAI},
  pages={12597--12604},
  year={2020}
}

Acknowledgments

Our code is based on open-reid and Adaptive-ReID, if you use our code, please also cite their paper.

@article{song2018unsupervised,
  title={Unsupervised domain adaptive re-identification: Theory and practice},
  author={Song, Liangchen and Wang, Cheng and Zhang, Lefei and Du, Bo and Zhang, Qian and Huang, Chang and Wang, Xinggang},
  journal={arXiv preprint arXiv:1807.11334},
  year={2018}
}

Resouces:

  1. Pretrained Models:

all pre-adapted models are named by the following formula:

ada{src}2{tgt}.pth

where "src" and "tgt" are the initial letter of source and target dataset's name, i.e., M for Market1501, D for Duke and C for CUHK03.

Baidu NetDisk, Password: 9aba

Google Drive

  1. Results on MSMT17(MS), Duke(D) and Market(M).

Results

About

code for AAAI 2020 paper "ACT"


Languages

Language:Python 100.0%