IsaacGuan / 3D-VAE

A variational autoencoder for volumetric shape generation

3D-VAE

This is the tf.keras implementation of the volumetric variational autoencoder (VAE) described in the paper "Generative and Discriminative Voxel Modeling with Convolutional Neural Networks".

Preparing the Data

Some experimental shapes from the ModelNet10 dataset are saved in the datasets folder. The model consumes volumetric shapes compressed in the TAR file format. For details about the structure and preparation of the TAR files, please refer to voxnet.

Training

python train.py

Testing

python test.py