The 3D Gaussian reconstruction method based on forward-rendered rasterization

This paper proposes the 3D Gaussian reconstruction method based on forward-rendered rasterization, aimed at addressing the high training and rendering costs associated with existing neural radiance field (NeRF) methods, which hinder real-time rendering with high visual quality. Unlike time-consuming neural network training approaches, this method introduces a novel 3D Gaussian scene representation combined with a real-time differentiable renderer, significantly accelerating scene optimization and novel view synthesis. By fitting the scene with 3D Gaussian distribution and using tile-based rasterization for rendering, this method achieves high-quality, real-time (≥30fps) novel view synthesis at 1080p resolution. Experimental results demonstrate that this approach achieves visual quality comparable to, and in some cases surpassing, the state-of-the-art Mip-NeRF360, while matching the training time of the fastest current methods. The proposed method successfully enables real-time rendering of scenes captured from multiple photographs.