GeoGaussian: Geometry-aware Gaussian Splatting for Scene Rendering

TL;DR

Gaussian Splatting can degrade scene geometry in low-textured regions, hindering novel-view synthesis. GeoGaussian addresses this by using surface-aligned thin Gaussians, a tangent-space densification strategy, and explicit geometry constraints to preserve geometry and appearance. The method encodes geometry meaning in Gaussian parameters and enforces coplanarity with neighbors to stabilize reconstruction and rendering. Across Replica, TUM RGB-D, and ICL-NUIM, GeoGaussian achieves state-of-the-art novel-view rendering and geometric reconstruction, especially under sparse training views, highlighting its practical impact for geometry-preserving 3D Gaussian representations.

Abstract

During the Gaussian Splatting optimization process, the scene's geometry can gradually deteriorate if its structure is not deliberately preserved, especially in non-textured regions such as walls, ceilings, and furniture surfaces. This degradation significantly affects the rendering quality of novel views that deviate significantly from the viewpoints in the training data. To mitigate this issue, we propose a novel approach called GeoGaussian. Based on the smoothly connected areas observed from point clouds, this method introduces a novel pipeline to initialize thin Gaussians aligned with the surfaces, where the characteristic can be transferred to new generations through a carefully designed densification strategy. Finally, the pipeline ensures that the scene's geometry and texture are maintained through constrained optimization processes with explicit geometry constraints. Benefiting from the proposed architecture, the generative ability of 3D Gaussians is enhanced, especially in structured regions. Our proposed pipeline achieves state-of-the-art performance in novel view synthesis and geometric reconstruction, as evaluated qualitatively and quantitatively on public datasets.

Figures (15)

• Figure 1: Comparisons of novel view rendering and 3D Gaussian model on the Replica Datasets. As highlighted in the second row, the proposed method shows a very clear boundary between two low-textured walls, but 3DGS has blurring issues since the geometry of its 3D Gaussian model is not accurate in this area.
• Figure 2: Geometry-aware strategies of our GeoGaussian. In smoothly connected areas, the parameterization of thin Gaussians contains clear geometry meanings in the mean vector and covariance matrix. Furthermore, the densification operation for these thin Gaussians encourages the new generations to lie in the tangent space established by the position and normal vectors of the original Gaussian. Finally, these thin Gaussians, measured by a training view, are used to establish smooth constraints with photometric constraints in the optimization process.
• Figure 3: Comparisons of novel view rendering on public datasets. At some challenging viewpoints having bigger differences in translation and orientation motions compared with training views, 3DGS and LightGS have issues with photorealistic rendering. (d) shows the training view closest to the rendered one.
• Figure 4: Statistics of the number of Gaussians in sequences of Replica. To make the comparison compact, more values are illustrated in Appendix.
• Figure 5: Rendering performance in training and evaluation using TUM RGB-D (a), Replica (b), and ICL-NUIM (c) datasets.