3D-HGS: 3D Half-Gaussian Splatting

Haolin Li,Jinyang Liu,Mario Sznaier,Octavia Camps

2024-06-14

Abstract:Photo-realistic 3D Reconstruction is a fundamental problem in 3D computer vision. This domain has seen considerable advancements owing to the advent of recent neural rendering techniques. These techniques predominantly aim to focus on learning volumetric representations of 3D scenes and refining these representations via loss functions derived from rendering. Among these, 3D Gaussian Splatting (3D-GS) has emerged as a significant method, surpassing Neural Radiance Fields (NeRFs). 3D-GS uses parameterized 3D Gaussians for modeling both spatial locations and color information, combined with a tile-based fast rendering technique. Despite its superior rendering performance and speed, the use of 3D Gaussian kernels has inherent limitations in accurately representing discontinuous functions, notably at edges and corners for shape discontinuities, and across varying textures for color discontinuities. To address this problem, we propose to employ 3D Half-Gaussian (3D-HGS) kernels, which can be used as a plug-and-play kernel. Our experiments demonstrate their capability to improve the performance of current 3D-GS related methods and achieve state-of-the-art rendering performance on various datasets without compromising rendering speed.

Computer Vision and Pattern Recognition,Graphics

What problem does this paper attempt to address?

The paper aims to address the issue of discontinuous function representation in 3D scene reconstruction, particularly the shape and color discontinuities at object edges and in texture-rich areas. Current methods, such as 3D Gaussian Splatting (3D-GS), perform well in terms of rendering performance and speed but have inherent limitations when dealing with these discontinuities. To this end, the authors propose a new method—3D Half-Gaussian Splatting (3D-HGS), which improves existing techniques by splitting each 3D Gaussian kernel into two half-Gaussian kernels with different opacity values. This approach allows for more accurate capture of high-frequency information while retaining the advantages of the original 3D Gaussian kernel. Experimental results show that 3D-HGS outperforms existing 3D-GS-related methods in rendering performance across multiple datasets and achieves state-of-the-art levels without sacrificing rendering speed. Additionally, this method, as a plug-and-play kernel, can be easily applied to existing methods based on 3D Gaussian kernels.