Research map / 3D vision / Depth and stereo
GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GeoSVR, a sparse-voxel (SVRaster-based) surface reconstruction framework that uses monocular depth with a voxel-level uncertainty weighting plus voxel dropout and surface regularizers to achieve accurate, complete meshes.
Problem
Surface reconstruction from multi-view images with radiance fields is limited by Gaussian-based representations that depend on sparse point-cloud initialization; sparse voxel representations lack a structure prior and suffer geometric distortion without external constraints.
Approach
Builds on SVRaster's sparse voxels and adds a level-aware voxel geometric uncertainty that reweights a patch-wise monocular depth loss per pixel, plus Voxel Dropout (random subsampling of voxels during multi-view regularization), Surface Rectification, and a Scaling Penalty to sharpen voxel-based surfaces.
Key results
- On DTU, GeoSVR achieves mean Chamfer distance 0.47, the best among compared methods (PGSR 0.52, GeoNeuS 0.51).
- On Tanks and Temples, GeoSVR achieves mean F1 0.56, versus 0.52 for PGSR and 0.50 for Neuralangelo.
- On Mip-NeRF 360, GeoSVR reaches 24.83 outdoor PSNR (best) and 30.46 indoor PSNR, with competitive rendering quality; ablation on TnT raises F1 from 0.397 (SVRaster base) to 0.560.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GeoSVR instead of PGSR | surface reconstruction · DTU | Chamfer distance: 0.52 -> 0.47 | Helped |
| use GeoSVR instead of PGSR | surface reconstruction · Tanks and Temples | F1-score: 0.52 -> 0.56 | Helped |
| add sparse points (ablation) | SVRaster (Base) · Tanks and Temples | F1-score: 0.397 -> 0.382 | Hurt |
| add inverse-depth supervision (ablation) | SVRaster (Base) · Tanks and Temples | F1-score: 0.397 -> 0.398 | No clear effect |
| add patch-wise depth loss (ablation) | SVRaster (Base) · Tanks and Temples | F1-score: 0.397 -> 0.449 | Helped |
| add multi-view regularization (ablation) | SVRaster with patch-wise depth · Tanks and Temples | F1-score: 0.449 -> 0.538 | Helped |
| add voxel dropout during multi-view regularization (ablation) | SVRaster with patch-wise depth and multi-view regularization · Tanks and Temples | F1-score: 0.538 -> 0.546 | Helped |
| add surface rectification (ablation) | GeoSVR without surface rectification · DTU | cf-dist: 0.477 -> 0.468 | Helped |
| add scaling penalty (ablation) | GeoSVR with surface rectification · Tanks and Temples | F1-score: 0.549 -> 0.552 | Helped |
| add voxel-uncertainty weighting to monocular depth (ablation) | GeoSVR with DepthAnythingV2 · Tanks and Temples | F1-score: 0.552 -> 0.560 | Helped |
| remove scene constraint (ablation) | GeoSVR · Tanks and Temples | F1-score: 0.560 -> 0.523 | Hurt |
| add patch-wise depth supervision (ablation) | PGSR · Tanks and Temples | F1-score: 0.527 -> 0.538 | Helped |
Surprising findings
- Vanilla SVRaster without scene constraint produces heavily distorted geometry, and 3DGS-based methods gain little from monocular depth constraints, unlike the sparse voxel method.
Limitations
- Struggles with reflective regions, textureless areas, and transparent surfaces.
- Rendering quality slightly drops when geometry is forced to be accurate.
- Multi-view regularization training is slow due to inefficient implementation.
Experimental setup
Datasets: DTU Tanks and Temples Mip-NeRF 360
Benchmarks: DTU Tanks and Temples Mip-NeRF 360
Metrics: Chamfer distance F1-score PSNR SSIM LPIPS training time FPS
Baselines: VolSDF NeuS Neuralangelo GeoNeuS MonoSDF 2DGS GOF SVRaster GS2Mesh VCR-GauS MonoGSDF PGSR
Builds on: SVRaster PGSR Depth Anything V2 3D Gaussian Splatting MonoSDF
Closest papers by meaning
- DySurface: Consistent 4D Surface Reconstruction via Bridging Explicit Gaussians and Implicit Functions (NeurIPS 2026 · similarity 0.90)
- GVKF: Gaussian Voxel Kernel Functions for Highly Efficient Surface Reconstruction in Open Scenes (NeurIPS 2024 · similarity 0.89)
- Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering (CVPR 2025 · similarity 0.89)
- Learning with Unreliability: Fast Few-shot Voxel Radiance Fields with Relative Geometric Consistency (CVPR 2024 · similarity 0.89)
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · similarity 0.88)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.