Skip to content

imagednn/Down-to-the-Last-Detail-Virtual-Try-on-with-Detail-Carving

 
 

Repository files navigation

Down to the Last Detail: Virtual Try-on with Detail Carving

Code for virtual try-on with high-fidelity details. The code was developed and tested with Pytorch0.4.1.

Virtual try-on results

Getting Started

Installation

  • Clone this repo
git clone https://github.com/AIprogrammer/Detailed-virtual-try-on.git. 
cd Detailed-virtual-try-on
  • Download our pretrained models from Google Drive, and put them in "./pretrained_checkpoint".

Demo

  • We provide a demo model, as well as some samples in "./dataset/images". Triplets including source image, target pose, target cloth is provided in the "./demo/demo.txt".
  • Quick testing and checking results in "./demo/forward/0.jpg" by running
sh demo.sh

Training

Download the dataset

  • Download the MPV dataset from Image-based Multi-pose Virtual Try On and put the dataset under "./dataset/images/".
  • Select postive perspective images, create dataset split file 'data_pair.txt', and put it under "./dataset/".

Dataset preprocessing

  • Pose keypoints. Use the Openpose, and put the keypoints file in "./dataset/pose_coco".
  • Semantic parsing. Use the CIHP_PGN, and put the parsing results in "./dataset/parse_cihp".
  • Cloth mask. Use the "GrabCut" method for the cloth mask, and put the mask in "./dataset/cloth_mask".

Coarse-to-fine training

  • Download the VGG19 pretrained checkpoint
cd vgg_model/
wget https://download.pytorch.org/models/vgg19-dcbb9e9d.pth
  • Set different configuration based on the "config.py". Then run
sh train.sh

Citation

If you find this code helpful, please cite our paper:

@inproceedings{detail2019,
  title={Down to the Last Detail: Virtual Try-on with Detail Carving},
  author={Wang, Jiahang and Zhang, Wei and Weizhong, Liu and Mei, Tao},
  booktitle = {arXiv:1912.06324},
  year={2019}
}

About

Virtural try-on under arbitrary poses

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.4%
  • Shell 0.6%