Improved Adversarial Systems for 3D Object Generation and Reconstruction

TL;DR

This paper introduces a Wasserstein GAN-based method for improved 3D object generation and reconstruction, enhancing training stability and accuracy in modeling complex 3D shapes from various data sources.

Contribution

It extends prior GAN approaches by incorporating Wasserstein distance with gradient penalty, enabling better training and more accurate 3D shape generation and reconstruction.

Findings

Achieves significant quantitative improvements over existing methods.

Successfully reconstructs 3D shapes from 2D images.

Performs shape completion from occluded 2.5D scans.

Abstract

This paper describes a new approach for training generative adversarial networks (GAN) to understand the detailed 3D shape of objects. While GANs have been used in this domain previously, they are notoriously hard to train, especially for the complex joint data distribution over 3D objects of many categories and orientations. Our method extends previous work by employing the Wasserstein distance normalized with gradient penalization as a training objective. This enables improved generation from the joint object shape distribution. Our system can also reconstruct 3D shape from 2D images and perform shape completion from occluded 2.5D range scans. We achieve notable quantitative improvements in comparison to existing baselines