Domain Adaptive Person Re-Identification via Coupling Optimization
Codes of our paper "Domain Adaptive Person Re-Identification via Coupling Optimization" ACM MM 2020. If you find this useful, please kindly cite our paper:
@inbook{liu2020dimglo,
author = {Liu, Xiaobin and Zhang, Shiliang},
title = {Domain Adaptive Person Re-Identification via Coupling Optimization},
year = {2020},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3394171.3413904},
pages = {547–555},
numpages = {9}
}
(Sample code uses Market-1501 as source dataset and DukeMTMC-reID as target dataset.)
Performance
Market->Duke: Rank1 0.755, mAP 0.575 Duke->Market: Rank1 0.870, mAP 0.643 Market->MSMT17: Rank1 0.487, mAP 0.201 Duke->MSMT17: Rank1 0.565, mAP 0.244
Dataset
-
Market-1501 [BaiduYun] [GoogleDriver] CamStyle (generated by CycleGAN) [GoogleDriver] [BaiduYun] (password: 6bu4)
-
DukeMTMC-reID [BaiduYun] (password: bhbh) [GoogleDriver] CamStyle (generated by CycleGAN) [GoogleDriver] [BaiduYun] (password: 6bu4)
-
MSMT17 + CamStyle (generated by StarGAN) [BaiduYun] (password: 6bu4) [GoogleDriver] We reformulate the structure of MSMT17 the same as Market-1501.
Images generated by GAN are provided by Zhun Zhong
Framework
pytorch >= 1.1.0
Change path to datasets:
Market-1501 (Source):
Change the path of Market-1501 to where your dataset are at Line 81 in train.py.
Duke (Target):
Change the path of DukeMTMC-reID to where your dataset are at Line 20 in reid/compute_memory_bank.py, Line 103 and Line 108 in reid/compute_map.py, Line 106, 107 and 108 in reid/DukeDataProvider.py.
The trained model with Market-1501 as source and Duke as target can be downloaded at here, this model achieves 0.761 and 0.583 in Rank1 and mAP accuracy, respectively.
Run the code:
Run python train.py in terminal
Contact me
If you have any question, please feel free to contact me: Xiaobin Liu