Awesome

person_generation_spt

Unsupervised Person Image Generation with Semantic Parsing Transformation <br>(CVPR 2019, oral).

Sijie Song, Wei Zhang, Jiaying Liu, Tao Mei

Project page: http://39.96.165.147/Projects/SijieSong_cvpr19/CVPR19_ssj.html

Check out our paper and supplementary here.

Prerequisites

• Python 2 (Python 3 should also work, but needs some modification)
• Pytorch >= 0.4.0
• opencv-python
• Numpy
• Scipy
• Pandas
• Skimage

Getting started

A demo model is given for appearance generation. We provide some samples in "./imgs", the parsing maps are in "./parsing".

• Clone this repo:

git clone https://github.com/SijieSong/person_generation_spt.git

cd person_generation_spt

• Download pre-trained models from Google Drive or Baidu Yun, put ./demo_model under ./checkpoints

• Quick testing (modify the gpu_id in ./scripts/test_demo.sh if needed)

bash ./scripts/test_demo.sh

• Check the results in ./results/demo_test (source image | target pose (ground truth) | output)

  <img src='./results/demo_test/2_A.jpg_2_B.jpg.png' width=400 alt="Example"/>

• Testing a new image:

  You can test a new image with pre-defined parsing files (see the example in ./parsing). The id for each attribute label is defined as below: 0-background, 1-face, 2-hair, 3-upperclothes, 4-pants, 5-skirt, 6-leftArm, 7-rightArm, 8-leftLeg, 9-rightLeg.

Citation

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

@inproceedings{song2019unsupervised,
  title={Unsupervised Person Image Generation with Semantic Parsing Transformation},
  author={Song, Sijie and Zhang, Wei and Liu, Jiaying and Mei, Tao},
  booktitle = {Proc.~IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

Related projects

• CycleGAN and pix2pix in PyTorch
• pose-gan
• LIP-SSL

Contact

Sijie Song ssj940920 AT pku.edu.cn