VoteHMR

the official implementation of ACM MM 2021 paper <VoteHMR:Occlusion-Aware Voting Network for Robust 3D Human Mesh Recovery from Partial Point Clouds>

Introduction

This is a release of our paper <VoteHMR:Occlusion-Aware Voting Network for Robust 3D Human Mesh Recovery from Partial Point Clouds>
Authors: Guanze Liu, Yu Rong, Lu Sheng*
[arxiv]

Part of the code is inspired by DCT-ICCV and VoteNet

Citation

@inproceedings{liu2021votehmr,
title={VoteHMR: Occlusion-Aware Voting Network for Robust 3D Human Mesh Recovery from Partial Point Clouds},
author={Liu, Guanze and Rong, Yu and Sheng, Lu},
booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
pages={955--964},
year={2021}
}

Installation

The code is tested under the following environment

• pytorch 1.1.0
• python 3.7.10
• cuda 10.1
• pointnet2 (https://github.com/facebookresearch/votenet/tree/main/pointnet2)

  Data Preparation

  Download Human Models

  the required SMPL models are uploaded to BaiduYUN [extract code: ame2] for download

  Set up Blender

You need to download Blender and install scipy package to run the first part of the code. The provided code was tested with Blender2.78, which is shipped with its own python executable as well as distutils package. Therefore, it is sufficient to do the following:

• Install pip

  /blenderpath/2.78/python/bin/python3.5m get-pip.py

• Install scipy

  /blenderpath/2.78/python/bin/python3.5m pip install scipy

• Install numpy

  /blenderpath/2.78/python/bin/python3.5m pip install numpy

Notice that get-pip.py is downloaded from pip. Replace the blenderpath with your own and set BLENDER_PATH.

Set up OpenEXR

In order to read rendered depth images and segm images, the OpenEXR bindings for PYTHON is required.
Set the openexr_py2_path in src/datasets/config.copy as your OpenEXR path.

Prepare Training Data

The Synthetic Dataset SURREAL and DFAUST can be downloaded from existing repositary.
We provide the shell script to generate partial point cloud from provided synthetic data.
You should set the corresponding data_path in the following shell scripts, including surreal_data_path, surreal_save_path, dfaust_save_path, as well as the tmp_path, output_path in src/datasets/config.copy

sh scripts/generate_training_data.sh

Prepare Testing Data

We also provide the shell script to generate testing data sh scripts/generate_testing_data.sh

Training

We provide the shell scripts to train with VoteHMR

• If you wish to train on single gpu, run the following shell script sh scripts/train.sh
• If you wish to train on multiple gpu, run the following shell script sh scripts/train_dist.sh

Set the surreal_save_path as the path to processed surreal data.

Evaluation

We provide the shell scripts for evaluation, set chechpoints as your model path dir.

sh scripts/test.sh

Fine Tuning on Real Data

If you wish to fine tune on real data, run
sh scripts/weakly_supervised.sh

For Online Visdom Support

Before training starts, to visualize the training results and loss curve in real-time, please run python -m visdom.server 8098 and click the URL https//localhost:8098