yxgeee / MMT-plus

Solution for VisDA 2020

• Technique Report: arXiv or GitHub
• Video Introduction: YouTube or bilibili

Requirements

• Python 3
• PyTorch >= 1.1

Installation

• Install visda library

git clone https://github.com/yxgeee/MMT-plus.git
cd MMT-plus
python setup.py develop

• Install NVIDIA/apex library (optional, for mixed precision training)

git clone https://github.com/NVIDIA/apex.git
cd apex
python setup.py install --cuda_ext --cpp_ext

Active it by adding --fp16 in the training commands.

Prepare Datasets

personX_sda can be downloaded from [Google Drive], while the others can be downloaded from [Simon4Yan/VisDA2020].

examples
├── data
│   ├── index_validation_gallery.txt  
│   ├── index_validation_query.txt  
│   ├── index_test_gallery.txt  
│   ├── index_test_query.txt     
│   ├── personX
│   ├── personX_sda
│   ├── target_training
│   ├── target_test
└── └── target_validation

Testing

The trained models for our submission can be downloaded from:

• Re-ID model:
  • ResNeSt-50 [Google Drive]
  • ResNeSt-101 [Google Drive]
  • DenseNet-169-IBN [Google Drive]
  • ResNeXt-101-IBN [Google Drive]
• Camera model:
  • ResNeSt-50 [Google Drive]

Test a single model

CUDA_VISIBLE_DEVICES=0 ./scripts/test.sh $ARCH_REID $PATH_REID $ARCH_CAMERA $PATH_CAMERA

If you want to test a model without domain-specific BN (e.g. pre-trained model), you need to remove --dsbn from scripts/test.sh.

Model ensembling and testing

CUDA_VISIBLE_DEVICES=0 ./scripts/test_ensemble.sh

Please make sure the model path in scripts/test_ensemble.sh is correct before testing.

Top-3 results on the leaderboard

Training

Stage I: Structured Domain Adaptation (SDA)

You could directly download the generated images from [Google Drive], or you could use the following scripts to train your own SDA model.

• Train SDA:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./scripts/train_sda.sh $ARCH $SOUECE_MODEL $TARGET_MODEL

• Generate source-to-target images:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./scripts/test_sda.sh $PATH_OF_GENERATOR

Stage II: Pre-training

CUDA_VISIBLE_DEVICES=0,1,2,3 ./scripts/pretrain.sh personx_sda $ARCH 1
CUDA_VISIBLE_DEVICES=0,1,2,3 ./scripts/pretrain.sh personx_sda $ARCH 2

Stage III: Improved Mutual Mean-Teaching (MMT+)

CUDA_VISIBLE_DEVICES=0,1,2,3 ./scripts/train_mmt_dbscan.sh $ARCH

Post-processing: Camera Classification Training

CUDA_VISIBLE_DEVICES=0 ./scripts/camera.sh $ARCH

The trained camera model will be used in the inference stage.

Citation

If you find this code useful for your research, please consider cite:

@misc{ge2020improved,
    title={Improved Mutual Mean-Teaching for Unsupervised Domain Adaptive Re-ID},
    author={Yixiao Ge and Shijie Yu and Dapeng Chen},
    year={2020},
    eprint={2008.10313},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

@inproceedings{ge2020mutual,
  title={Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification},
  author={Yixiao Ge and Dapeng Chen and Hongsheng Li},
  booktitle={International Conference on Learning Representations},
  year={2020},
  url={https://openreview.net/forum?id=rJlnOhVYPS}
}

@misc{ge2020structured,
    title={Structured Domain Adaptation with Online Relation Regularization for Unsupervised Person Re-ID},
    author={Yixiao Ge and Feng Zhu and Rui Zhao and Hongsheng Li},
    year={2020},
    eprint={2003.06650},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Acknowledgement

This code is mainly based on MMT.