Paper Image

Structured 3D Gaussians for Generative Modeling

Published on:

28 March 2024

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Bowen Zhang,

Yiji Cheng,

Jiaolong Yang,

Chunyu Wang,

Feng Zhao,

Yansong Tang,

Dong Chen,

Baining Guo


Key Details

Proposes GaussianCube, a structured 3D Gaussian representation for generative modeling

Performs densification-constrained fitting to obtain fixed number of Gaussians

Rearranges Gaussians into voxel grid via Optimal Transport

Achieves efficient and high-fidelity 3D reconstruction

Enables high-quality 3D generation with simple 3D U-Net architecture

AI generated summary

Structured 3D Gaussians for Generative Modeling

This paper proposes GaussianCube, a novel 3D representation that structures Gaussian Splatting into a voxel grid using Optimal Transport. This allows high-quality 3D reconstruction and efficient rendering, while making the representation amenable to generative modeling with standard 3D convolutional networks. Experiments show state-of-the-art unconditional and conditional generation on ShapeNet and OmniObject3D datasets.

Answers from this paper


No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up