MUST-GAN

Code | paper

The Pytorch implementation of our CVPR2021 paper "MUST-GAN: Multi-level Statistics Transfer for Self-driven Person Image Generation".

Tianxiang Ma, Bo Peng, Wei Wang, Jing Dong,

CRIPAC,NLPR,CASIA & University of Chinese Academy of Sciences.

---

Test results of our model under self-supervised training:

Pose transfer

Clothes style transfer

Requirement

• python3
• pytorch 1.1.0
• numpy
• scipy
• scikit-image
• pillow
• pandas
• tqdm
• dominate
• visdom

Getting Started

Installation

• Clone this repo:

git clone https://github.com/TianxiangMa/MUST-GAN.git
cd MUST-GAN

Data Preperation

We train and test our model on Deepfashion dataset. Especially, we utilize High-Res Images in the In-shop Clothes Retrieval Benchmark.

Download this dataset and unzip (You will need to ask for password.) it, then put the folder img_highres under the ./datasets directory. Download train/test split list, which are used by a lot of methods, and put them under ./datasets directory. - Run the following code to split train/test dataset.

python tool/generate_fashion_datasets.py

Download source-target paired images list, as same as the list used by many previous work. Becouse our method can self-supervised training, we do not need the fashion-resize-pairs-train.csv, you can download train_images_lst.csv for training.

Download train/test keypoints annotation files and semantic segmentation files.

Put all the above files into the ./datastes folder.

• Run the following code to generate pose map and pose connection map.

python tool/generate_pose_map.py
python tool/generate_pose_connection_map.py

Download vgg pretrained model for training, and put it into ./datasets folder.

Test

Download our pretrained model, and put it into ./check_points/MUST-GAN/ folder.

• Run the following code, and set the parameters as your need.

bash scripts/test.sh

Train

• Run the following code, and set the parameters as your need.

bash scripts/train.sh

Citation

If you use this code for your research, please cite our paper:

@InProceedings{Ma_2021_CVPR,
    author    = {Ma, Tianxiang and Peng, Bo and Wang, Wei and Dong, Jing},
    title     = {MUST-GAN: Multi-Level Statistics Transfer for Self-Driven Person Image Generation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {13622-13631}
}

Acknowledgments

Our code is based on PATN and ADGAN, thanks for their great work.