Abstract

Geometric priors are often used to enhance 3D reconstruction. With manysmartphones featuring low-resolution depth sensors and the prevalence ofoff-the-shelf monocular geometry estimators, incorporating geometric priors asregularization signals has become common in 3D vision tasks. However, theaccuracy of depth estimates from mobile devices is typically poor for highlydetailed geometry, and monocular estimators often suffer from poor multi-viewconsistency and precision. In this work, we propose an approach for jointsurface depth and normal refinement of Gaussian Splatting methods for accurate3D reconstruction of indoor scenes. We develop supervision strategies thatadaptively filters low-quality depth and normal estimates by comparing theconsistency of the priors during optimization. We mitigate regularization inregions where prior estimates have high uncertainty or ambiguities. Ourfiltering strategy and optimization design demonstrate significant improvementsin both mesh estimation and novel-view synthesis for both 3D and 2D GaussianSplatting-based methods on challenging indoor room datasets. Furthermore, weexplore the use of alternative meshing strategies for finer geometryextraction. We develop a scale-aware meshing strategy inspired by TSDF andoctree-based isosurface extraction, which recovers finer details from Gaussianmodels compared to other commonly used open-source meshing tools. Our code isreleased in https://xuqianren.github.io/ags_mesh_website/.