Research map / 3D vision / Gaussian splatting
GVKF: Gaussian Voxel Kernel Functions for Highly Efficient Surface Reconstruction in Open Scenes
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 GVKF, a voxel-grid-based implicit 3DGS representation with kernel-regression opacity field and a logistic opacity-to-SDF mapping, enabling efficient surface reconstruction and novel view synthesis in open outdoor scenes.
Problem
Efficient and high-fidelity 3D surface reconstruction and novel view synthesis in large open scenes, where NeRF is slow and 3DGS/2DGS use huge numbers of Gaussian primitives with high memory use and rough surfaces in sparse areas.
Approach
Manage 3D Gaussians via a sparse voxel grid with MLP-decoded attributes, and replace discrete Gaussian opacities in alpha blending with Gaussian kernel functions so a continuous opacity field is fit along rays; a logistic mapping then converts the opacity CDF to an SDF for mesh extraction.
Key results
- On Waymo, GVKF reaches 30.24 PSNR vs 28.51 for 2DGS, with 30 MB storage and 14 GB GPU memory versus 238 MB and 23 GB for 2DGS.
- On Tanks and Temples, GVKF achieves mean F1 0.36, above explicit baselines (3DGS 0.09, 2DGS 0.30, SuGaR 0.19) but below Neuralangelo (0.50).
- On Mip-NeRF 360, GVKF reports PSNR 25.47 and LPIPS 0.240 on outdoor scenes, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GVKF instead of 2DGS | Waymo Open Scene reconstruction · Waymo Open Scene | PSNR: 28.51 -> 30.24 | Helped |
| Use GVKF instead of NeuS | Waymo Open Scene reconstruction · Waymo Open Scene | Chamfer distance: 0.76 -> 1.57 | Hurt |
| Use voxel-based GVKF representation | Waymo Open Scene reconstruction · Waymo Open Scene | Storage: 230 -> 30 MB | Helped |
| Use GVKF instead of 2DGS | Novel view synthesis · Mip-NeRF 360 outdoor | PSNR: 24.33 -> 25.47 | Helped |
| Use GVKF instead of GOF | Novel view synthesis · Mip-NeRF 360 | PSNR: 24.53 -> 25.47 | Helped |
| Use GVKF instead of GOF | Surface reconstruction · Tanks and Temples | F1: 0.46 -> 0.36 | Hurt |
| Reduce voxel size from 0.1 to 0.01 (ablation) | GVKF on Waymo · Waymo | PSNR: 29.34 -> 30.24 | Helped |
| Reduce voxel size from 0.01 to 0.001 (ablation) | GVKF on Waymo · Waymo | PSNR: 30.24 -> 30.29 | No clear effect |
| Use voxel size 1 (ablation) | GVKF on Waymo · Waymo | PSNR | Hurt |
| Add voxel representation and its MLPs (ablation) | GVKF · Tanks and Temples | PSNR: 23.60 (-2.71) -> 26.31 | Helped |
| Add voxel representation and its MLPs (ablation) | GVKF · Tanks and Temples | F1: 0.39 (+0.03) -> 0.36 | Hurt |
| Use the opacity-to-SDF mapping instead of a linear assumption (ablation) | GVKF · Tanks and Temples | F1: 0.30 (-0.06) -> 0.36 | Helped |
Surprising findings
- Against concurrent GOF, GVKF has lower Tanks and Temples mean F1 (0.36 vs 0.46) while using far less storage (68 MB vs 649 MB on Mip-NeRF 360).
Limitations
- Local line-of-sight SDF fitting leaves regions not covered by training views with uneven surfaces.
- Struggles with dynamic objects, sky being misrepresented as a surface, and decoupling of near and far views.
- Geometric precision lags GOF on Tanks and Temples and meshing of the opacity-to-SDF mapping needs numerical root solving that adds time.
Experimental setup
Datasets: Waymo Open Dataset Tanks and Temples Mip-NeRF 360
Benchmarks: Waymo Open Dataset Tanks and Temples Mip-NeRF 360
Metrics: PSNR SSIM LPIPS Chamfer distance F1 score FPS training time GPU memory storage
Baselines: NeuS F2-NeRF StreetSurf 3DGS SuGaR 2DGS Geo-NeuS Neuralangelo BakedSDF GOF
Builds on: 3D Gaussian Splatting Gaussian Opacity Fields Scaffold-GS Octree-GS 2D Gaussian Splatting
Closest papers by meaning
- GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction (NeurIPS 2025 · similarity 0.89)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.89)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.89)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · similarity 0.89)
- GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- GSDF: 3DGS Meets SDF for Improved Neural Rendering and Reconstruction (NeurIPS 2024 · similarity 0.89)
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.