Human Mesh Recovery from Monocular Images via a Skeleton-disentangled Representation

Sun Yu, Ye Yun, Liu Wu, Gao Wenpeng, Fu YiLi, Mei Tao

Accepted to ICCV 2019 https://arxiv.org/abs/1908.07172 Paper Link

Internet Video Demo

More Demos on YouTube

Requirements

Python 3.6+
Pytorch tested on 0.4.1/1.0/1.2 versions
PyTorch implementation of the Neural 3D Mesh Renderer for visualization

Installation

pip install -r requirements.txt

Demo

Simply go into DSD-SATN/src/, and run

sh run.sh

The results are saved in DSD-SATN/resources/results.

Re-implementation

Prepare model and data.

Step 1) Download the pre-trained models and statistical model from google drive. Unzip them under the project dir (e.g. DSD-SATN/trained_model, DSD-SATN/model)

Step 2) Download the processed annotations (moshed parameters, extracted DSD features of images in Human3.6M, and 2D pose estimations from AlphaPose) of Human3.6M and 3DPW dataset. Unzip them and set their location in data_set_path of src/config.py like

data_set_path = {
    'h36m':'PATH/TO/H36M',
    ...
    'pw3d':'/PATH/TO/3DPW',}

Step 3) Apply for the datasets from official Human3.6M and 3DPW. Especially, pre-processing the input images of Human3.6M dataset. Extracting 1 frame from every 5 frame of video in Human3.6M dataset. Set the path of extracted images as {H36m_dir}/images and name each image as the format (Sn_action name_camera id_frame number.jpg) shown in h36m/h36m_test.txt (e.g. S11_Discussion 2_2_149.jpg).

Re-implement the evaluation results on Human3.6M and 3DPW datasets.

# Evaluating single-frame DSD network on Human3.6M dataset
CUDA_VISIBLE_DEVICES=0 python3 test.py --gpu=0 --dataset=h36m --tab=single_h36m --eval

# Evaluating entire network (DSD-SATN) on Human3.6M dataset
CUDA_VISIBLE_DEVICES=0 python3 test.py --gpu=0 --dataset=h36m --tab=video_h36m --eval --video --eval-with-single-frame-network

# Evaluating single-frame DSD network on 3DPW dataset
CUDA_VISIBLE_DEVICES=0 python3 test.py --gpu=0 --tab=single_3dpw --eval --eval-pw3d

# Evaluating entire network (DSD-SATN) on 3DPW dataset
CUDA_VISIBLE_DEVICES=0 python3 test.py --gpu=0 --tab=video_3dpw --eval --eval-pw3d --video --eval-with-single-frame-network

Saving results & Visualization

Additionally, if you want to save the results, please add some options:

--save-obj # saving obj file of 3D human body mesh.
--save-smpl-params # saving smpl parameters.
--visual-all # saving all visual rendering results.

For example, saving all results of single-frame DSD network on Human3.6M dataset, just type

CUDA_VISIBLE_DEVICES=0 python3 test.py --gpu=0 --dataset=h36m --tab=single_h36m --eval --save-obj --save-smpl-params --visual-all

TODO List

Citation

If you use this code for your research, please consider citing:

@InProceedings{sun2019dsd-satn,
title = {Human Mesh Recovery from Monocular Images via a Skeleton-disentangled Representation},
author = {Sun, Yu and Ye, Yun and Liu, Wu and Gao, Wenpeng and Fu, YiLi and Mei, Tao},
booktitle = {IEEE International Conference on Computer Vision, ICCV},
year = {2019}
}

Acknowledgement

we refer to pytorch_hmr for training code. The fast rendering module is brought from face3d. The transformer module is brought from transformer-pytorch.

xiaobaozi1996 / DSD-SATN