Awesome

Structure-Aware Flow Generation for Human Body Reshaping (CVPR 2022)

Jianqiang Ren, Yuan Yao, Biwen Lei, Miaomiao Cui, Xuansong Xie

DAMO Academy, Alibaba Group, Hangzhou, China

News

(2022-09-28) The pretrained model and code are available now.

Overview

We propose a novel end-to-end structure-aware flow generation framework for human body reshaping(FBBR), which can achieve favorable and controllable results for high-resolution images efficiently. The BR-5K is the first large-scale dataset for body reshaping, it consists of 5,000 high-quality individual portrait photos at 2K resolution collected from Unsplash.

BR5K Dataset

Considering that the misuse of the dataset may lead to ethical concerns, as recommended by AC, we will review the application to access the datasets. To be able to download the BR5K database, please download, sign the agreement form, and then use your work e-mail(e.g., xx@xx.edu.cn, xx@your_company.com) to send the form to (jianqiang.rjq@alibaba-inc.com).

Getting Started

Install Requriements

Python >= 3.6
torch >= 1.2.0
numba

Models

Download the pose estimator model body_pose_model.pth from here.
Download the reshaping model body_reshape_model.pth from google drive or 百度网盘 (key:17e6).
Put body_pose_model.pth and body_reshape_model.pth into the models folder.

Run the Demo

python test.py --config config/test_demo_setting.toml

the results will be in the test_cases_output directory.

Quantitative Evaluaton

In this repository, we take a new pretrained model of openpose and achieve better quantitative performance than reported in our paper.

Method	SSIM	PSNR	LPIPS
Baseline	0.8339	24.4916	0.0823
Paper	0.8354	24.7924	0.0777
This Repository	0.8394	25.5801	0.0684

Citation

If our work is useful for your research, please consider citing:

@inproceedings{ren2022structure,
title={Structure-Aware Flow Generation for Human Body Reshaping},
author={Ren, Jianqiang and Yao, Yuan and Lei, Biwen and Cui, Miaomiao and Xie, Xuansong},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={7754--7763},
year={2022}
}

Acknowledgement

We express gratitudes to openpose and pytorch-openpose, as we benefit a lot from both their papers and codes.