jihoonkim888 / 3d-gan

Fine-scale Synthetic Shape Generation with Upsampling Network

This repository contains the code implementation of the 3D-GAN and the up-convolutional network that are used to generate synthetic samples and upsample low-resolution samples, respectively.

Root

• train_3D_GAN.py: trains 3D-GAN used to generate synthetic shape samples
• train_upscaler.py: trains the proposed up-convolutonal network to upsample low-resolution shape samples
• main.py: generates low-resolution synthetic samples, upsample them and output 3D voxel plots of the original and upsampled shape samples
• main_128.py: generates high-resolution synthetic samples and output their 3D voxel plots
• main_real.py: plots real low and high-resolution shape samples used to train the 3D-GAN and the up-convolutional network
• main_real_ups.py: upsamples real low-resolution shape samples and output their 3D voxel plots

Subfolders

The repository contains two subfolders, obj_to_voxel and src.

obj_to_voxel

• binvox: program that converts a mesh (.obj) file in into an occupancy grid (.binvox)
• voxelize_obj.py: converts all the mesh files in a folder into .binvox files. it should be able to convert meshes in any formats that binvox supports.
• Dockerfile: A script to build a docker image that can run binvox and voxelize_obj.py

src

• binvox_rw.py: A Python script that reads .binvox files as arrays (numpy)
• GAN.py: The source code of the 3D-GAN used to generate shape samples
• upscaler.py: The source code of the up-convolutional network used to upsample shape samples

You will need an NVIDIA GPU that supports CUDA with at least 4GB of RAM to run the low-res 3D-GAN or the up-convolutional network, and 8GB of VRAM to run the high-resolution 3D-GAN.

The docker image used to train and test can be found here. For voxelisation of mesh files, please use the Dockerfile in the subfolder.

References

• binvox https://www.patrickmin.com/binvox/
• binvox_rw.py https://github.com/dimatura/binvox-rw-py