Generative and Discriminative Voxel Modeling with Convolutional Neural Networks

TL;DR

This paper investigates voxel-based 3D data representations, introducing new neural network methods for shape modeling and object classification, achieving significant improvements on benchmark datasets.

Contribution

It presents novel training methods for voxel-based variational autoencoders, a user interface for latent space exploration, and a deep CNN architecture for object classification.

Findings

51.5% relative improvement on ModelNet benchmark

Effective voxel-based shape modeling and classification

Addressed challenges specific to voxel representations

Abstract

When working with three-dimensional data, choice of representation is key. We explore voxel-based models, and present evidence for the viability of voxellated representations in applications including shape modeling and object classification. Our key contributions are methods for training voxel-based variational autoencoders, a user interface for exploring the latent space learned by the autoencoder, and a deep convolutional neural network architecture for object classification. We address challenges unique to voxel-based representations, and empirically evaluate our models on the ModelNet benchmark, where we demonstrate a 51.5% relative improvement in the state of the art for object classification.