PaperlayerView exact revision on arXiv

Source-linked AI summary

GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction

Jiahe Li, Jiawei Zhang, Youmin Zhang, Xiao Bai, Jin Zheng, Xiaohan Yu, Lin Gu

arXiv:2509.18090v2cs.CV

TL;DR

Radiance-field surface reconstruction remains constrained by Gaussian-based representational bottlenecks and sparse voxels’ missing scene constraints and local refinement. GeoSVR addresses these issues with an explicit voxel framework, uncertainty-aware monocular depth constraints, and sparse-voxel surface regularization. Experiments report superior performance across challenging scenarios in geometric accuracy, detail preservation, completeness, and efficiency.

  • Problem

    Gaussian-based reconstruction faces representational bottlenecks, while sparse voxels lack strong scene constraints and have highly local surface refinement.

  • Method

    GeoSVR combines explicit sparse voxels with Voxel-Uncertainty Depth Constraint and Sparse Voxel Surface Regularization.

  • Results

    GeoSVR achieves superior surface reconstruction performance across DTU, Tanks and Temples, and Mip-NeRF 360 while preserving detail, completeness, and computational efficiency.

  • Takeaways & Limitations

    Sparse voxels can support accurate, detailed, and complete surface reconstruction when their uncertainty and local geometric consistency are explicitly addressed.

  • Takeaways & Limitations

    Serious reflections, textureless areas, and transparent surfaces can still cause suboptimal geometry because of photometric inconsistency and limited ray-tracing representation.

Abstract

from arXiv · show

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.

1 Introduction

GeoSVR explores sparse voxels as an explicit alternative to Gaussian-based reconstruction, targeting complete, clear, and efficient surface modeling. It addresses sparse-voxel distortion and locality through uncertainty-aware depth constraints and surface regularization.

  • 1 Introduction: Existing Gaussian-based methods are limited by incomplete point-cloud initialization and ambiguous Gaussian geometry, motivating an alternative representation.Inaccurate or uncovered regions hinder refinement, while Gaussian primitives lack clearly defined edges.
  • 1 Introduction: GeoSVR uses fully covered, clearly bounded sparse voxels to support complete geometry learning and more identifiable details than Gaussians or smooth neural fields.The framework starts from SVRaster and optimizes explicit sparse voxels with RGB images.
  • 1 Introduction: Sparse-voxel optimization suffers from geometric distortion because it lacks strong structure priors and because tiny voxels have highly local influence during surface formation.These challenges motivate external depth constraints and broader geometric consistency regularization.
  • 1 Introduction: GeoSVR introduces Voxel-Uncertainty Depth Constraint to use monocular depth cues while avoiding quality degradation through voxel uncertainty evaluation.The contribution is framed as an effective and robust scene constraint for accuracy-sensitive reconstruction.
  • 1 Introduction: Sparse Voxel Surface Regularization enlarges global consistency for tiny voxels and facilitates sharp, accurate voxel-based surface formation.Together with the explicit voxel framework, it targets delicate and complete geometry learning.
  • 1 Introduction: Experiments on DTU, Tanks and Temples, and Mip-NeRF 360 report superior reconstruction performance across challenging scenarios while preserving detail, completeness, and computational efficiency.The reported advantages span geometric accuracy, detail preservation, and reconstruction completeness.

2 Related Works

Related work spans implicit neural fields, hybrid grid representations, and explicit Gaussian-based surface reconstruction. GeoSVR is presented within this progression as a sparse-voxel alternative emphasizing explicit geometric clarity and efficient refinement.

  • 2 Related Works: NeRF-based methods learn implicit fields through differentiable rendering but are computationally expensive, while later neural grids, plane decompositions, and sparse voxels improve representation efficiency.These hybrid representations commonly assume uniformly scaled grids, limiting quality and scalability.
  • 2 Related Works: 3D Gaussian Splatting balances fast, high-quality reconstruction through differentiable rasterized splatting, but intersecting Gaussians can produce view-inconsistent rendering and rely on sparse point clouds.Gaussian-based surface methods subsequently integrate SDFs or improved representations to advance surface quality.
  • 2 Related Works: GeoSVR uses sparse voxels with rasterization and level-of-detail refinement, providing complete coverage and clearer geometric boundaries than Gaussian or smooth neural representations.Its overview combines uncertainty-aware depth constraints with voxel-based regularization for surface refinement.

3 Method

GeoSVR uses sparse voxels to preserve complete scene coverage and geometric clarity, addressing their weak native constraints and highly local surface refinement. It combines voxel-aware monocular-depth supervision with surface regularization to guide convergence and improve geometric accuracy.

  • 3.1 Preliminaries: GeoSVR represents scenes with sparse voxels that provide complete coverage and clearly bounded geometry, unlike point-cloud-dependent explicit approaches.SVRaster organizes sparse voxels in an adaptive octree and refines them through rasterization, pruning, and subdivision.
  • 3.1 Preliminaries: Sparse voxels still suffer geometry distortion because they lack strong structure priors, while tiny voxels affect only nearby neighbors during surface formation.These limitations motivate both an essential scene constraint and a broader geometric-consistency regularizer.
  • 3.2 Voxel Geometric Uncertainty for Scene Constraint: Monocular depth supplies dense, efficient constraints matching the spatially complete voxels, but inaccurate depth requires adaptive reliance to avoid degrading reconstruction quality.The paper identifies the absence of learned-geometry confidence estimation as a key obstacle to using this cue effectively.
  • 3.2 Voxel Geometric Uncertainty for Scene Constraint: Voxel geometric uncertainty increases for coarse, low-level voxels and critical-density regions, reflecting limited geometric accuracy, texture constraints, or view coverage.The uncertainty combines an Octree-level base term with voxel density and is used to identify where external depth supervision is most needed.
  • 3.2 Voxel Geometric Uncertainty for Scene Constraint: The Voxel-Uncertainty Depth Constraint converts rendered level-aware uncertainty into pixel weights, reducing monocular-depth influence where native photometric evidence is reliable and increasing it for ambiguous voxels.It applies these weights to a patch-wise global-local depth loss for scale alignment and robust geometry learning.
  • 3.3 Sparse Voxel Surface Regularization: Surface refinement must correct rendering-density misalignment, because trilinear voxel interactions can shift the highest rendering weight away from the highest-density surface location.This motivates Surface Rectification and further voxel-level regularization for accurate surface formation.

4 Experiments

GeoSVR is evaluated on DTU, Tanks and Temples, and Mip-NeRF 360 against implicit, explicit, and geometry-cue baselines. It achieves strong geometric accuracy, completeness, detail preservation, and efficiency, while ablations attribute gains to its scene and voxel regularization designs.

  • Surface Reconstruction: GeoSVR outperforms all baselines in DTU overall accuracy and achieves the best Tanks and Temples F1-score, including against SDF, 3DGS, and external-cue methods.The comparisons use Chamfer distance and F1-score on DTU and Tanks and Temples.
  • Surface Reconstruction: GeoSVR obtains both the best accuracy and completeness while retaining fast training comparable to 3DGS-based methods.Its densely covered, prior-free sparse voxels also handle reflective and insufficient-coverage regions where 3DGS methods are limited.
  • Appearance Reconstruction: GeoSVR delivers competitive novel-view synthesis on Mip-NeRF 360 among surface-reconstruction and NVS-specific baselines.Surface quality is not quantitatively evaluated there because the dataset lacks geometry ground truth.
  • Scene Constraint: The Voxel-Uncertainty Depth Constraint identifies uncertain regions and refines geometry while preserving already well-reconstructed parts, improving a reconstruction with F1 0.552.Patch-wise depth loss helps geometry more than sparse-depth or inverse-loss alternatives in the ablation.
  • Multi-view and Voxel Regularization: Voxel Dropout further improves multi-view consistency beyond explicit multi-view regularization and a patch-warping reference method with monocular depth.Surface Rectification and Scaling Penalty additionally improve voxel-based surface formation quantitatively and qualitatively.

5 Conclusion

GeoSVR extends sparse voxels into an efficient framework for accurate, detailed, and complete surface reconstruction. It uses voxel uncertainty for robust external scene constraints and addresses voxel-based surface refinement, while identifying voxel globality as a future direction.

  • 5 Conclusion: GeoSVR explores sparse voxels for accurate, detailed, and complete surface reconstruction with high efficiency.The framework analyzes voxel uncertainty and voxel-based surface refinement as central reconstruction challenges.
  • 5 Conclusion: Voxel uncertainty enables effective and robust scene constraints from external cues by distinguishing confidence in learned geometry.The conclusion identifies enhanced voxel globality as a future direction for varying lights and textureless regions.

A Ablation Study

The ablations show that scene constraints are necessary to prevent distorted sparse-voxel geometry, while voxel dropout and surface rectification improve local consistency and surface accuracy. These effects are examined across challenging regions and DTU and Tanks and Temples evaluations.

  • A.1 Scene Constraint: Removing scene constraints causes obvious, heavily distorted geometry and drags overall performance despite competitive accuracy in partial regions.The proposed scene constraint resolves this drawback, including in challenging regions where 3DGS-based methods remain limited by initialization.
  • A.1 Scene Constraint: Monocular depth improves GeoSVR substantially on challenging regions, and the Voxel-Uncertainty Depth Constraint remains effective with DepthPro despite weaker outdoor robustness.Compared with PGSR, GeoSVR better exploits monocular depth because it is less dependent on initial SfM points.
  • A.2 Voxel Dropout: Voxel Dropout improves geometry regularization by relieving local traps caused when tiny voxels receive gradients only from nearby regions.Qualitative comparisons show it addresses redundant geometric structures that regularization alone only marginally penalizes.
  • A.3 Surface Rectification: Surface Rectification acts as fine-grained regularization for tiny voxels with critical density changes, with its effect especially visible on DTU.Additional DTU ablations quantify its contribution to highly accurate surface reconstruction.

B Derivation of Voxel Geometric Uncertainty

The derivation models voxel geometric uncertainty from information retention, voxel scale, local geometric information, and an adjustment coefficient, then approximates it for stable weighting.

  • B Derivation of Voxel Geometric Uncertainty: A voxel’s desired geometric information is modeled as local information density multiplied by its voxel volume, while each voxel has a scale-independent maximum representation capacity Gmax.The voxel side length is related to octree level by vs = ws × 2^-l.
  • B Derivation of Voxel Geometric Uncertainty: The maximum information-retention ratio ηmax provides an upper bound for underfitted regions where sparse voxels cannot capture unlimited-resolution geometry.This limitation arises from finite computational resources.
  • B Derivation of Voxel Geometric Uncertainty: Voxel geometric uncertainty is built from the inverse information-retention ratio and a local target adjustment coefficient g(v), so weights reflect voxel-specific information loss.The construction is intended to prevent the uncertainty weight from shrinking to zero when ideal voxel capacity meets the target.
  • B Derivation of Voxel Geometric Uncertainty: The raw uncertainty approximation can become numerically unstable at large octree levels because it contains powers and exponent-related scale factors.The derivation therefore removes the overall cubic power, replaces 1/2^l with 1/l, and introduces a coefficient to control value variation.
  • B Derivation of Voxel Geometric Uncertainty: The final uncertainty formulation uses a level-dependent base term multiplied by a density-based saturation factor, Ugeom(v) = Ubase(l) · (1 − exp(−vgeo)).The base term is defined as Ubase(l) = ws / β(l + l0).

D.1 Datasets

Evaluation uses DTU, Tanks and Temples, and Mip-NeRF 360 datasets, with established baselines and official or reproduced results under matched processing settings.

  • D.1 Datasets: The evaluation covers 15 DTU scans, 6 high-quality Tanks and Temples scenes, and all 9 Mip-NeRF 360 scenes.DTU uses half-resolution training images and preprocessing from 2DGS through COMLAP; TnT uses publicly accessible ground truth.
  • D.1 Datasets: Baselines include implicit SDF methods such as NeuS and Neuralangelo, plus explicit methods including 2DGS, PGSR, MonoGSDF, and SVRaster.SVRaster uses sparse voxels, most other explicit baselines use 3DGS, and MonoGSDF uses a hybrid SDF.
  • D.1 Datasets: Several baselines incorporate external geometry cues, including monocular depth, monocular normals, stereo surfaces, or multi-view confidence checks.MonoGSDF uses DepthAnythingV2 monocular depth, the same depth model used by GeoSVR.
  • D.1 Datasets: Qualitative comparisons prioritize official checkpoints or results, while unavailable scores are reproduced using official code or substituted with reported results from another published method.For Geo-NeuS on Tanks and Temples, Neuralangelo’s reported results are used for comparison.

D.3 Metrics

Metrics are selected by dataset and aligned with prior work, while qualitative mesh visualizations cover Tanks and Temples, Mip-NeRF 360, and DTU.

  • D.3 Metrics: Chamfer distance measures DTU accuracy, F1-Score measures overall Tanks and Temples quality, and prior 3DGS metrics are retained for Mip-NeRF 360.The evaluation uses corresponding established toolkits and dependency versions for fairness.
  • D.3 Metrics: Average rendering speed is reported separately in Table 8 to assess inference efficiency across datasets.The supplied passage identifies the table’s scope but does not provide its numerical entries.
  • D.3 Metrics: Mesh visualizations are provided for Tanks and Temples and Mip-NeRF 360, including qualitative comparisons and reconstructed surfaces.Additional DTU visualizations include meshes and vertex-colored meshes.

F Efficiency Analysis

The efficiency analysis finds that proposed components generally preserve efficiency, with multi-view regularization contributing most to increased training time.

  • F Efficiency Analysis: All ablation components maintain high inference FPS, memory efficiency, and voxel-count efficiency, except for added training-time cost from multi-view regularization.The authors attribute that cost mainly to less efficient code implementation and plan to address it in future work.
  • F Efficiency Analysis: The proposed components seldom increase GPU memory under a similar voxel count, supported by efficient constraint selection and restrained voxel-level regularization.The passage also attributes the result to efficient coding implementation.

G Additional Visualization Results

GeoSVR reconstructs high-quality meshes across DTU, Tanks and Temples, and Mip-NeRF 360, preserving vivid object detail in complex scenes. These capabilities support practical use but warrant caution about malicious applications and privacy.

  • G Additional Visualization Results: GeoSVR reconstructs high-quality meshes across complex scenes in Tanks and Temples, Mip-NeRF 360, and DTU.The visualizations include reconstructed scenes, objects, and colored renderings.
  • G Additional Visualization Results: The reconstructed objects preserve accurate detail, supporting the method’s practical value in real-world applications.
  • Accurate reconstructions may be misused and may raise privacy concerns when derived from real-world data.The authors report no direct negative societal impact discovered so far, while urging caution during use.

I Discussion

GeoSVR extends sparse-voxel reconstruction toward accurate, complete, and detailed geometry while maintaining efficiency. Its performance remains constrained in reflective, textureless, and transparent regions, motivating future work on ray tracing, voxel globality, and transparency.

  • I Discussion: GeoSVR achieves state-of-the-art accuracy, completeness, and detail preservation through an explicit voxel-based representation.The discussion frames this as extending the potential of voxel-based recovery beyond Gaussian Splatting-focused approaches.
  • I Discussion: Serious reflections, textureless areas, and transparent surfaces remain principal limitations that often produce suboptimal geometry.Photometric inconsistency and limited ray-tracing representation capability contribute to these failures.
  • I Discussion: Accurate geometry can slightly reduce rendering quality when real-world images lack ideal photometric multi-view consistency.Reflective and transparent materials can make distorted geometry a better local solution for matching appearance.
  • I Discussion: Future improvements could combine more efficient ray tracing, improved voxel globality, and dedicated transparency solutions.
Loading 2509.18090v2…