# Generative and Discriminative Voxel Modeling with Convolutional Neural Networks

**Authors:** Andrew Brock, Theodore Lim, J. M. Ritchie, Nick Weston  
**arXiv:** [1608.04236](https://arxiv.org/abs/1608.04236)  
**Date:** 2016-08-15  
**Summary page:** https://www.summarizepaper.com/en/paper/1608.04236v2/

## Key points

- The importance of choosing the right representation when working with three-dimensional data is explored
- Voxel-based models are supported as effective representations in various applications, including shape modeling and object classification
- Methods for training voxel-based variational autoencoders are developed to learn a latent space representation
- A user interface is introduced for exploration of the learned latent space
- A deep convolutional neural network architecture specifically designed for object classification using voxel-based representations is proposed
- Solutions to challenges related to voxel-based representations are provided
- Empirical evaluation on the ModelNet benchmark dataset shows a significant improvement in object classification performance compared to state-of-the-art methods (51.5% relative improvement)
- Voxel-based models are shown to be viable and effective in handling three-dimensional data

## Summary

The paper titled "Generative and Discriminative Voxel Modeling with Convolutional Neural Networks" explores the importance of choosing the right representation when working with three-dimensional data. The authors focus on voxel-based models and provide evidence to support the use of voxellated representations in various applications, including shape modeling and object classification. The main contributions of this research include the development of methods for training voxel-based variational autoencoders, which are used to learn a latent space representation. Additionally, the authors introduce a user interface that allows for exploration of this learned latent space. They also propose a deep convolutional neural network architecture specifically designed for object classification using voxel-based representations. One of the key challenges addressed in this work is related to voxel-based representations. The authors provide solutions to overcome these challenges and empirically evaluate their models on the ModelNet benchmark dataset. The results demonstrate a significant improvement in object classification performance, achieving a relative improvement of 51.5% compared to state-of-the-art methods. Overall, this paper highlights the viability and effectiveness of voxel-based models in handling three-dimensional data. The proposed methods for training variational autoencoders, exploring latent spaces, and classifying objects contribute to advancing the field of shape modeling and object classification using voxel-based representations.