GausSurf: Geometry-Guided 3D Gaussian Splatting for Surface Reconstruction

TL;DR

GausSurf tackles the challenge of high-quality surface reconstruction from posed RGB images by embedding geometry priors into a 3D Gaussian Splatting framework. It couples patch-match based MVS guidance in texture-rich regions with normal priors in texture-less regions, arranged in an iterative scheme that mutualistically refines Gaussian parameters and depth/normal maps. The approach achieves state-of-the-art speed among explicit Gaussian-based methods while delivering competitive or superior surface quality on DTU and Tanks and Temples, and maintains strong novel-view synthesis performance. This geometry-guided, data-efficient strategy broadens the practical applicability of Gaussian splatting for accurate surface reconstruction and faster training, with potential for future real-time SLAM extensions.

Abstract

3D Gaussian Splatting has achieved impressive performance in novel view synthesis with real-time rendering capabilities. However, reconstructing high-quality surfaces with fine details using 3D Gaussians remains a challenging task. In this work, we introduce GausSurf, a novel approach to high-quality surface reconstruction by employing geometry guidance from multi-view consistency in texture-rich areas and normal priors in texture-less areas of a scene. We observe that a scene can be mainly divided into two primary regions: 1) texture-rich and 2) texture-less areas. To enforce multi-view consistency at texture-rich areas, we enhance the reconstruction quality by incorporating a traditional patch-match based Multi-View Stereo (MVS) approach to guide the geometry optimization in an iterative scheme. This scheme allows for mutual reinforcement between the optimization of Gaussians and patch-match refinement, which significantly improves the reconstruction results and accelerates the training process. Meanwhile, for the texture-less areas, we leverage normal priors from a pre-trained normal estimation model to guide optimization. Extensive experiments on the DTU and Tanks and Temples datasets demonstrate that our method surpasses state-of-the-art methods in terms of reconstruction quality and computation time.

Figures (7)

• Figure 1: Given a set of posed RGB images (a), our method reconstructs high-quality surfaces (f) with greater efficiency compared to existing Gaussian Splatting-based methods (b-e). In each subcaption (b-f), the first row indicates the reconstruction method, while the second row shows the average reconstruction time and chamfer distance on the DTU dataset, respectively.
• Figure 2: Visualization of geometric priors. (Scene 97 in the DTU dataset aanaes2016dtu) (a) Reference image; (b) Refined depth map using patch-match, where the background deep purple color indicates removed unreliable pixels; (d) Refined normal map using patch-match; (d) Estimated normal prior generated by StableNormal ye2024stablenormal.
• Figure 3: Patch-match based Geometry Guidance. Given the rendered depth/normal from Gaussians, we leverage patch-matching to refine the depth and normal for future Gaussian optimization, via propagation, random perturbation and multi-view geometric check.
• Figure 4: Visualization of geometric priors. (Caterpillar in TnT dataset knapitsch2017tnt) (a) Reference image; (b) Refined depth map using patch-match, where the background purple color indicates removed unreliable pixels; (c) Estimated normal prior generated by StableNormal ye2024stablenormal.
• Figure 5: Qualitative comparisons of surface reconstruction on the DTU dataset. Our method can reconstruct more smooth and detailed surface with high efficiency.