GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction

TL;DR

GeoSVR tackles accurate surface reconstruction from images by adopting an explicit sparse-voxel representation (SVRaster) with an Octree backbone, avoiding dependence on Gaussian primitives and imperfect MVG point clouds. It introduces a Voxel-Uncertainty Depth Constraint to adaptively fuse monocular depth cues based on per-voxel geometric uncertainty, paired with Sparse Voxel Surface Regularization that enforces global geometric consistency and sharp surface formation through voxel dropout, Surface Rectification, and a Scaling Penalty. The method achieves state-of-the-art geometric accuracy, detail preservation, and completeness across DTU, Tanks and Temples, and Mip-NeRF 360 whileMaintaining competitive training speed. These results demonstrate the potential of explicit sparse voxel representations for robust, efficient 3D surface reconstruction, with promising directions for integrating stronger global constraints and handling challenging reflective or textureless regions.

Abstract

Reconstructing accurate surfaces with radiance fields has achieved remarkable progress in recent years. However, prevailing approaches, primarily based on Gaussian Splatting, are increasingly constrained by representational bottlenecks. In this paper, we introduce GeoSVR, an explicit voxel-based framework that explores and extends the under-investigated potential of sparse voxels for achieving accurate, detailed, and complete surface reconstruction. As strengths, sparse voxels support preserving the coverage completeness and geometric clarity, while corresponding challenges also arise from absent scene constraints and locality in surface refinement. To ensure correct scene convergence, we first propose a Voxel-Uncertainty Depth Constraint that maximizes the effect of monocular depth cues while presenting a voxel-oriented uncertainty to avoid quality degradation, enabling effective and robust scene constraints yet preserving highly accurate geometries. Subsequently, Sparse Voxel Surface Regularization is designed to enhance geometric consistency for tiny voxels and facilitate the voxel-based formation of sharp and accurate surfaces. Extensive experiments demonstrate our superior performance compared to existing methods across diverse challenging scenarios, excelling in geometric accuracy, detail preservation, and reconstruction completeness while maintaining high efficiency. Code is available at https://github.com/Fictionarry/GeoSVR.

Figures (14)

• Figure 1: Geometric Sparse-Voxel Reconstruction. Our method, abbreviated as GeoSVR, delivers high-quality surface reconstruction for intricate real-world scenes based on explicit sparse voxels. Our superiority is exhibited compared to the state-of-the-art approaches built upon Gaussian Splatting, which encounter rough, inaccurate, or incomplete recovery problems even with help from external estimators, excelling in delicate details capturing with high completeness and top-tier efficiency.
• Figure 2: Overview of GeoSVR. Our method starts from constantly initialized sparse voxels, optimized with RGB images. (a) To enforce correct scene convergence while avoiding accuracy degradation, we apply Voxel-Uncertainty Depth Constraint by evaluating geometric uncertainty to determine the degree of reliance on monocular depth cue. (b) Voxel Dropout is introduced to enlarge the global geometry consistency for tiny voxels during the explicit geometry regularization. (c) For fine-grained surface refinement, we align the voxel-level density field to the surfaces with Voxel Regularization, facilitating accurate and sharp surface formation.
• Figure 3: Illustration of Surface Rectification and the visualized process on voxels.
• Figure 4: Reconstructed Mesh Visualization on the DTU jensen2014dtu Dataset. Our GeoSVR achieves superior reconstruction both in accuracy and completeness, handling difficult regions well by geometry cue constraints while still preserving fine-grained details. Better visualized with zoom in.
• Figure 5: Reconstructed Mesh Visualization on the Tanks and Templesknapitsch2017tanksDataset. Our GeoSVR stands out by reconstructing accurate surfaces even for difficult scenes like complex buildings and weak texture regions, delivering intricate details as well as precise flats.