DRGSplat: Depth-Regularized 3D Gaussian Splatting

Daniel Barath, Keisuke Tateno, Marc Pollefeys, Federico Tombari

Keywords: gaussian splatting

Abstract: 3D Gaussian Splatting has rapidly become a leading method for photorealistic novel view synthesis. However, its geometric accuracy often lags behind its visual fidelity. Existing methods to improve geometry typically constrain the 3D Gaussians directly, compromising their volumetric nature. We introduce DRGSplat, a novel depth-regularization approach for 3D Gaussian Splatting that enhances geometric accuracy without direct modifications of the Gaussian primitives. DRGSplat regularizes the rendered depth maps during training with three key losses: a monocular depth loss enforcing global consistency, a surface normal loss refining local detail, and a new uncertainty-aware curvature loss that selectively penalizes high-gradient regions while avoiding the gradient instabilities common to direct curvature constraints. Experiments on standard benchmarks show that DRGSplat keeps the strong photometric quality of Gaussian Splatting while substantially improving geometric accuracy and outperforming state-of-the-art geometry-focused methods. On the ETH3D dataset, DRGSplat improves reconstruction accuracy, completeness, and F1 scores of 3DGS by 15, 25, and 17 percentage points, respectively. The source code will be made publicly available.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7179