GAGS: Gradient-Guided Adaptive Gaussian Splatting for Efficient and Geometry-Regularized Surface Reconstruction

Abstract. While 3D Gaussian Splatting (3DGS) has advanced surface reconstruction, existing implementations face critical challenges in memory efficiency and geometric fidelity. Current approaches like PGSR generate excessive Gaussian primitives due to uncontrolled densification, leading to redundant memory consumption while struggling with surface artifacts in boundary regions. This paper introduces GAGS (Gradient-Guided Adaptive Gaussian Splatting), a geometry-regularized framework that addresses these limitations through three technical contributions: photometric gradient-driven adaptive densification that strategically controls primitive subdivision using image gradient analysis, anisotropy-aware shape regularization for adaptive Gaussian scale constraint, and a dual regularization mechanism combining normal self-smoothing with depth-aware correction. Evaluations on the DTU dataset demonstrate the framework’s effectiveness in maintaining visual quality while significantly reducing redundant primitives— achieving an 83% reduction compared to PGSR baseline—with improved surface regularity in complex geometric regions. Project web: https://3241674469.github.io/GAGS-project/