Research map / 3D vision / NeRF and radiance fields
Geometry Field Splatting with Gaussian Surfels
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.
Splats a stochastic geometry field parameterized by Gaussian surfels with a near-exact differentiable volume splatting renderer and a distance-based color blending that makes the loss continuous, for opaque surface reconstruction from calibrated images.
Problem
Reconstructing accurate geometry of opaque surfaces from calibrated RGB images with Gaussian splatting, where existing density-based splatting relies on approximations (Taylor expansion, no self-occlusion, non-overlapping assumption) and produces a discontinuous loss landscape.
Approach
Parameterize the stochastic geometry field (not the volume density) with 2D Gaussian surfels, convert it to density with a closed-form footprint, render with refined volume splatting that keeps self-occlusion, and blend kernel colors by distance so the rendered color is continuous; replace SH colors with a latent-plus-MLP view-dependent color using reflection vectors for specular surfaces.
Key results
- On DTU (15 scans), Ours (SH) reaches average Chamfer Distance 0.63 and Ours (Latent) 0.58, versus 0.72 for 2DGS and 0.73 for GOF (as reported in the tables; 2DGS is listed at 0.72).
- On BlendedMVS scene-level, Ours (SH) achieves average Chamfer Distance 1.52 (x10^-1), best among compared methods, versus 1.69 for RaDe-GS and 1.71 for 2DGS.
- On BlendedMVS object-centric, Ours (Latent) averages 0.66 (x10^-2) versus 0.72 for RaDe-GS and 0.89 for 2DGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use geometry field splatting instead of original 2DGS splatting (ablation) | Geometry Field Splatting with Gaussian Surfels (SH) · DTU | Chamfer Distance: 0.70 -> 0.63 | Helped |
| use geometry field splatting instead of original 2DGS splatting (ablation) | Geometry Field Splatting with Gaussian Surfels (Latent) · DTU | Chamfer Distance: 0.72 -> 0.58 | Helped |
| add color propagation to remedy loss landscape defects (ablation) | Geometry Field Splatting with Gaussian Surfels (SH) · DTU | Chamfer Distance: 0.67 -> 0.63 | Helped |
| add color propagation to remedy loss landscape defects (ablation) | Geometry Field Splatting with Gaussian Surfels (Latent) · DTU | Chamfer Distance: 0.59 -> 0.58 | Helped |
| use per-ray sorting instead of global sorting (ablation) | Geometry Field Splatting with Gaussian Surfels (SH) · DTU | Chamfer Distance: 0.63 -> 0.61 | Helped |
| use per-ray sorting instead of global sorting (ablation) | Geometry Field Splatting with Gaussian Surfels (Latent) · DTU | Chamfer Distance: 0.58 -> 0.57 | Helped |
| use latent color representation instead of SH | Geometry Field Splatting with Gaussian Surfels · DTU | Chamfer Distance (Avg. ↓): 0.63 -> 0.58 | Helped |
| use latent color representation instead of SH | Geometry Field Splatting with Gaussian Surfels · BlendedMVS scene-level cases | Chamfer Distance (×10−1, Avg. ↓): 1.52 -> 1.71 | Hurt |
| use latent color representation instead of SH | Geometry Field Splatting with Gaussian Surfels · BlendedMVS object-centric cases | Chamfer Distance (×10−2, Avg. ↓): 0.67 -> 0.66 | Helped |
Surprising findings
- Using the latent color representation instead of SH is worse on BlendedMVS scene-level cases (1.71 vs 1.52), indicating the specular-oriented color design does not uniformly help.
Limitations
- Cannot handle transparent or semi-transparent objects due to the opaque assumption
- Fuzzy objects are not handled well due to the smoothness constraint
- Uses global sorting as an approximation to per-ray sorting
Experimental setup
Datasets: DTU BlendedMVS
Benchmarks: DTU BlendedMVS
Metrics: Chamfer Distance
Baselines: NeuS2 Neuralangelo Miller et al. (O.V.) NeUDF Chen et al. (F.D.S.) SuGaR 2DGS Dai et al. (G.S.) GOF RaDe-GS
Builds on: 2DGS volume splatting (Zwicker et al.) stochastic geometry field (Miller et al.) Gaussian splatting NeRF
Closest papers by meaning
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.90)
- SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction (NeurIPS 2025 · similarity 0.90)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference (NeurIPS 2025 · similarity 0.89)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.89)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · 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.