README.md

PFD：Pose-guided Feature Disentangling for Occluded Person Re-identification based on Transformer

This repo is the official implementation of "Pose-guided Feature Disentangling for Occluded Person Re-identification based on Transformer(PFD), Tao Wang, Hong Liu, Pinhao Song, Tianyu Guo& Wei Shi" in PyTorch.PFD-Net

Pipeline

Dependencies

• timm==0.3.2

• torch==1.6.0

• numpy==1.20.2

• yacs==0.1.8

• opencv_python==4.5.2.54

• torchvision==0.7.0

• Pillow==8.4.0

Installation

pip install -r requirements.txt

If you find some packages are missing, please install them manually.

Prepare Datasets

mkdir data

Please download the dataset, and then rename and unzip them under the data

• Occluded-Duke
• Occluede-REID
• Market-1501
• DukeMTMC-reID
• MSMT17

data
|--market1501
|
|--Occluded_Duke
|
|--Occluded_REID
|
|--MSMT17
|
|--dukemtmcreid

Prepare ViT Pre-trained and HRNet Pre-trained Models

mkdir data

The ViT Pre-trained model can be found in ViT_Base, The HRNet Pre-trained model can be found in HRNet, please download it and put in the './weights' dictory.

Training

We use One GeForce GTX 1080Ti GPU for Training Before train the model, please modify the parameters in config file, please refer to Arguments in TransReID

python occ_train.py --config_file {config_file path}
#examples
1. For Occluded-Duke:
python occ_train.py --config_file 'configs/OCC_Duke/skeleton_pfd.yml'
2. For Market-1501:
python occ_train.py --config_file 'configs/market1501/skeleton_pfd.yml'
3. For DUKEMTMC:
python occ_train.py --config_file 'configs/dukemtmcreid/skeleton_pfd.yml'
......

Test the model

First download the Occluded-Duke model:Occluded-Duke

To test on pretrained model on Occ-Duke: Modify the pre-trained model path (PRETRAIN_PATH:ViT_Base, POSE_WEIGHT:HRNet, WEIGHT:Occluded-Duke) in yml, and then run:

## OccDuke for example
python test.py --config_file 'configs/OCC_Duke/skeleton_pfd.yml'

Occluded-Duke Results

Model	Image Size	Rank-1	mAP
HOReID	256*128	55.1	43.8
PAT	256*128	64.5	53.6
TransReID	256*128	64.2	55.7
PFD	256*128	67.7	60.1
TransReID*	256*128	66.4	59.2
PFD*	256*128	69.5	61.8

$*$means the encoder is with a small step sliding-window setting

Occluded-REID Results

Model	Image Size	Rank-1	mAP
HOReID	256*128	80.3	70.2
PAT	256*128	81.6	72.1
PFD	256*128	79.8	81.3

Market-1501 Results

Model	Image Size	Rank-1	mAP
HOReID	256*128	80.3	70.2
PAT	256*128	95.4	88.0
TransReID	256*128	95.4	88.0
PFD	256*128	95.5	89.6

DukeMTMC Results

Model	Image Size	Rank-1	mAP
HOReID	256*128	86.9	75.6
PAT	256*128	88.8	78.2
TransReID	256*128	89.6	80.6
PFD	256*128	90.6	82.2
TransReID*	256*128	90.7	82.0
PFD*	256*128	91.2	83.2

Citation

If you find our work useful in your research, please consider citing this paper! (preprint version will be available soon)

Arxiv:

@misc{wang2021poseguided,
      title={Pose-guided Feature Disentangling for Occluded Person Re-identification Based on Transformer}, 
      author={Tao Wang and Hong Liu and Pinhao Song and Tianyu Guo and Wei Shi},
      year={2021},
      eprint={2112.02466},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

AAAI:

@inproceedings{wang2022pose,
  title={Pose-guided feature disentangling for occluded person re-identification based on transformer},
  author={Wang, Tao and Liu, Hong and Song, Pinhao and Guo, Tianyu and Shi, Wei},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={3},
  pages={2540--2549},
  year={2022}
}

Acknowledgement

Our code is extended from the following repositories. We thank the authors for releasing the codes.

• TransReID
• HRNet

License

This project is licensed under the terms of the MIT license.