Pseudo Label Rectification with Joint Camera Shift Adaptation and Outlier Progressive Recycling for Unsupervised Person Re-Identification [pdf]
The official repository for Pseudo Label Rectification With Joint Camera Shift Adaptation and Outlier Progressive Recycling for Unsupervised Person Re-Identification TITS'22.
We provide packaged python packages, which can be directly downloaded and unzipped to your server.
address: python3
# you can use this python package to run python scripts like:
/yourpath/reid/bin/python hello_world.py
# pip install something like:
/yourpath/reid/bin/python -m pip install numpy[INFO] This environment has packaged all the required dependencies. Please do not modify it at will！
Please make sure your dataset path is as follows：
## CCL
/YourPath/CASTOR/CCL/examples
├── data
│ ├── dukemtmcreid
│ │ └── DukeMTMC-reID
│ ├── market1501
│ │ └── Market-1501-v15.09.15
│ └── msmt17
│ └── MSMT17_V1
Download the datasets:
For privacy reasons, we don't have the copyright of the dataset. Please contact authors to get this dataset.
DukeMTMC-reID/
├── bounding_box_test
├── bounding_box_train
└── query
Market-1501-v15.09.15/
├── bounding_box_test
├── bounding_box_train
├── gt_bbox
├── gt_query
└── query
MSMT17_V1/
├── list_gallery.txt
├── list_query.txt
├── list_train.txt
├── list_val.txt
├── train
└── test
- When training with the backbone of IBN-ResNet, you need to download the ImageNet-pretrained model from this link and save it under the path of
CASTOR/CCL/examples/pretrained/andCASTOR/IDM/examples/pretrained/. - ImageNet-pretrained models for ResNet-50 will be automatically downloaded in the python script.
- You need download our trained camera classification model from this link and unzip it to
CASTOR/CCL/examples/pretrained/andCASTOR/IDM/examples/pretrained/
Please make sure your pretrain models path is as follows：
## CCL
/YourPath/CASTOR/CCL/examples/pretrained
├── resnet50_ibn_a.pth.tar (if you want to train IBN-ResNet)
├── resnet50-19c8e357.pth (it will be automatically downloaded by the python script)
├── camera_model/
│ ├── market1501
│ │ └── model_best.pth.tar
│ ├── dukemtmc
│ │ └── model_best.pth.tar
│ ├── msmt17
│ │ └── model_best.pth.tar
cd CASTOR/CCL/
# train command
sh scripts/market1501.sh
# test command refer to https://github.com/alibaba/cluster-contrast-reid
# View script to determine model save path!!!
sh scripts/test_market1501.sh| Method | mAP | Rank-1 | Rank-5 | Rank-10 | Download |
|---|---|---|---|---|---|
| CCL | 82.1 | 92.3 | 96.7 | 97.9 | link |
| CCL + CASTOR | 86.2 | 94.8 | 98.4 | 98.8 | model+log |
cd CASTOR/CCL/
# train command
sh scripts/dukemtmc.sh
# test command refer to https://github.com/alibaba/cluster-contrast-reid
# View script to determine model save path!!!
sh scripts/test_dukemtmc.sh| Method | mAP | Rank-1 | Rank-5 | Rank-10 | Download |
|---|---|---|---|---|---|
| CCL | 72.6 | 84.9 | 91.9 | 93.9 | link |
| CCL + CASTOR | 75.5 | 88.6 | 93.7 | 95.0 | model+log |
cd CASTOR/CCL/
# train command
sh scripts/msmt17.sh
# test command refer to https://github.com/alibaba/cluster-contrast-reid
# View script to determine model save path!!!
sh scripts/test_msmt17.sh| Method | mAP | Rank-1 | Rank-5 | Rank-10 | Download |
|---|---|---|---|---|---|
| CCL | 27.6 | 56.0 | 66.8 | 71.5 | link |
| CCL + CASTOR | 37.3 | 70.4 | 80.1 | 83.5 | model+log |
Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.
If you find this code useful for your research, please cite our paper
@ARTICLE{9967431,
author={Xu, Mingyuan and Guo, Haiyun and Jia, Yuheng and Dai, Zhitao and Wang, Jinqiao},
journal={IEEE Transactions on Intelligent Transportation Systems},
title={Pseudo Label Rectification With Joint Camera Shift Adaptation and Outlier Progressive Recycling for Unsupervised Person Re-Identification},
year={2022},
volume={},
number={},
pages={1-12},
doi={10.1109/TITS.2022.3224233}}
If you have any question, please feel free to contact us. E-mail: [email protected].