Awesome

Global-to-Local Generative Model for 3D Shapes

Introduction

Three-dimensional content creation has been a central research area in computer graphics for decades. The main challenge is to minimize manual intervention, while still allowing the creation of a variety of plausible 3D objects. In this work, we present a global-to-local generative model to synthesize 3D man-made shapes.

Global-to-Local Generative Model for 3D Shapes consist of two networks: a Global-to-Local GAN(G2LGAN) and a Part Refiner(PR).

Specifically, using a generative adversarial network training, with global and local discriminators, out G2LGAN generates 3D semantically segmented models. The PR is based on an auto-encoder architecture and its goal is to refine the individual semantic parts output from G2LGAN. For more details, please refer to our paper.

Usage

In our experiment, all the codes are tested under Tensorflow 1.2.1 GPU version and Python 2.7 on Ubuntu 16.04.

Train G2LGAN

To train a G2LGAN model to generate initial 3D shapes with semantic parts:

    cd G2LGAN
    python G2LGAN_train.py

To see all optional arguments for training:

    python G2LGAN_train.py -h
    

Generate Shapes from a pre-trained G2LGAN model

You can also sample some 3D shapes from pre-trained model:

    cd G2LGAN
    python G2LGAN_test.py 
    

Train Part Refiner

To train a Part Refiner model to refine the initial 3D shapes:

    cd Part_Refiner
    python train_part_refiner.py  
    

Use the pre-trained Part Refiner to refine the initial shape

Based on the pre-trained PR model, you can refine the initial generated shapes from G2LGAN by:

    cd Part_Refiner
    python test_part_refiner.py         
    

Visulizaton Tool

If you want to visualize the trainig data(.mat files) or generated shapes(combined by a .npy files), you can use the Visulizaton Tool we provide, just run:

    cd visualization_tool
    python application.py

You can manually choose the shape file on the tool, and take the visualization_tool/operation guide.pdf as reference.

License

Our code is released under MIT License (see LICENSE file for details).

Citation

Please cite the paper in your publications if it helps your research:

@article{G2L18,
title = {Global-to-Local Generative Model for 3D Shapes},
author = {Hao Wang and Nadav Schor and Ruizhen Hu and Haibin Huang and Daniel Cohen-Or and Hui Huang},
journal = {ACM Transactions on Graphics (Proc. SIGGRAPH ASIA)},
volume = {37},
number = {6},
pages = {214:1—214:10},  
year = {2018},
}