Abstract
Recently, 3D Gaussian Splatting (3DGS) has demonstrated impressive novel viewsynthesis results, while allowing the rendering of high-resolution images inreal-time. However, leveraging 3D Gaussians for surface reconstruction posessignificant challenges due to the explicit and disconnected nature of 3DGaussians. In this work, we present Gaussian Opacity Fields (GOF), a novelapproach for efficient, high-quality, and compact surface reconstruction inunbounded scenes. Our GOF is derived from ray-tracing-based volume rendering of3D Gaussians, enabling direct geometry extraction from 3D Gaussians byidentifying its levelset, without resorting to Poisson reconstruction or TSDFfusion as in previous work. We approximate the surface normal of Gaussians asthe normal of the ray-Gaussian intersection plane, enabling the application ofregularization that significantly enhances geometry. Furthermore, we develop anefficient geometry extraction method utilizing marching tetrahedra, where thetetrahedral grids are induced from 3D Gaussians and thus adapt to the scene'scomplexity. Our evaluations reveal that GOF surpasses existing 3DGS-basedmethods in surface reconstruction and novel view synthesis. Further, itcompares favorably to, or even outperforms, neural implicit methods in bothquality and speed.