Technical Explanation of 3D Gaussian Splatting

Abstract

Neural Radiance Fields (NeRF) has revolutionized novel-view synthesis of scenes captured through multiple photographs. Despite its remarkable visual quality, NeRF requires intensive computational resources for training and rendering due to the ray casting operations intertwined with neural network modules. Additionally, the definition of the 3D model as an implicit function of NeRF poses challenges in terms of editability. To overcome these limitations, 3D Gaussian Splatting (3DGS) has been proposed, enabling real-time rendering at Full-HD resolution using explicit 3D expression and efficient rasterization. This paper reviews the framework of 3DGS in comparison with NeRF and presents its results and applications.