Advancing Structured Priors for Sparse-Voxel Surface Reconstruction

TL;DR

This work presents SVRaster, a sparse-voxel surface reconstruction framework that combines the strengths of 3D Gaussian splats and sparse voxel rasterization via a geometry-guided Level-of-Detail voxel initialization and direct depth supervision. It introduces per-view SVOs, topology-aligned fusion, and an uncertainty-aware opacity model to produce a complete, high-fidelity initialization, followed by per-scene optimization with refined depth targets $D_t^{\star}$. The approach demonstrates faster convergence and improved geometric accuracy on benchmarks like DTU, outperforming both implicit and explicit baselines and showing robustness to foundation-model priors. The method enables crisp surface recovery with strong completeness and efficient mesh export, and points to future work in handling illumination, texture-poor regions, and scalable scenes with richer priors.

Abstract

Reconstructing accurate surfaces with radiance fields has progressed rapidly, yet two promising explicit representations, 3D Gaussian Splatting and sparse-voxel rasterization, exhibit complementary strengths and weaknesses. 3D Gaussian Splatting converges quickly and carries useful geometric priors, but surface fidelity is limited by its point-like parameterization. Sparse-voxel rasterization provides continuous opacity fields and crisp geometry, but its typical uniform dense-grid initialization slows convergence and underutilizes scene structure. We combine the advantages of both by introducing a voxel initialization method that places voxels at plausible locations and with appropriate levels of detail, yielding a strong starting point for per-scene optimization. To further enhance depth consistency without blurring edges, we propose refined depth geometry supervision that converts multi-view cues into direct per-ray depth regularization. Experiments on standard benchmarks demonstrate improvements over prior methods in geometric accuracy, better fine-structure recovery, and more complete surfaces, while maintaining fast convergence.

Figures (8)

• Figure 1: Our method initializes voxels at plausible regions with the associated levels of detail. With the geometry is refined by direct depth supervision, our SVRaster reconstruction better captures detailed 3D geometry. Our initialization-seeded design couples fast convergence with high geometric fidelity and strong completeness throughout training.
• Figure 2: Overview of Our Method. Given images $\{I_i\}$ with cameras $\{C_i\}$, our method reconstructs surfaces via per-scene training. First, a pretrained model estimates pseudo geometry priors (i.e., cameras $\{\tilde{C_i}\}$, depth $\{\tilde{D_i}\}$, and confidence $\{\tilde{\kappa_i}\}$), which are utilized to build per-image octrees $\{V_i\}$ with our level-of-detail unprojection. The resulting per-image octrees are fused into a single octree $V^{\mathrm{fusion}}$ (as an initialized SVO), along with the prediction of the opacity of each fused voxel grid. In the per-scene SVO optimization stage, we derive the refined depths for each camera view via advancing the geometry supervision (i.e., $\mathcal{L}_{\mathrm{refD}}$) from the rendered depths $\{D_i\}$ and the associated input images. This geometry-supervision loss, when combined with the original SVRaster training loss, improves the geometric accuracy of the reconstructed surface.
• Figure 3: Topology-Aligned Octree Fusion. To integrate sparse octrees derived from different camera views, the final resolution of each voxel is determined by the per-view finest voxels occupying the same spatial cell. For example, the rightmost voxel of octree A at level 2 is decomposed into multiple level-3 voxels to align the structure of octree B. The decomposed voxels inherit properties (e.g. color) from their parent voxel. To obtain the final fused octree, we average the colors of the corresponding voxels across per-view octrees as the blended voxels (denoted by the purple voxels.
• Figure 4: Rendered Normals on DTU Dataset. Our method preserves the local details while GeoSVR produces overly smoothed results.
• Figure 5: Convergence on DTU Dataset. Our method stabilizes training and accelerates convergence compared to the baseline. Note that, without specific geometry regularization, SVRaster sun2025sparse is not expected to outperform GeoSVR in terms of convergence.