Research map / 3D vision / Gaussian splatting
GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting
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.
Learns unsigned distance functions for open-surface reconstruction from multi-view images by overfitting 2D Gaussian planes with projection, self-supervision along Gaussian normals, and depth/normal constraints, without ray tracing.
Problem
Reconstructing open surfaces with sharp boundaries from multi-view images by learning unsigned distance functions, which is hard to do with discrete explicit 3D Gaussian representations and complex UDF gradients near the surface.
Approach
Overfit thin 2D Gaussian planes onto the zero level set of a jointly learned UDF (via stop-gradient projection), then supervise UDF values near the surface by sampling offsets along Gaussian normals and far from the surface with a gradient-based Chamfer projection loss.
Key results
- On DeepFashion3D, achieves mean Chamfer Distance 1.60 (×10^-3), best among compared methods including VRPrior (1.71) and 2S-UDF (1.98), with 1.6h training versus 7.8h to 9.2h for UDF baselines.
- On DTU, average Chamfer Distance 0.68 vs. 0.71 for VRPrior and 0.74 for GOF, the best among compared baselines.
- Ablation on DTU shows removing the self-supervision near loss raises CD from 0.68 to 0.77, the largest drop among removed terms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GaussianUDF instead of VRPrior | open-surface reconstruction · DF3D | Chamfer Distance (×10−3): 1.71 -> 1.60 | Helped |
| use GaussianUDF instead of VRPrior | open-surface reconstruction · DF3D | training time: 9.2h -> 1.6h | Helped |
| use GaussianUDF instead of VRPrior | surface reconstruction · DTU | Chamfer Distance: 0.71 -> 0.68 | Helped |
| add Lnear self-supervision to Lfar (ablation) | GaussianUDF · DTU | Chamfer Distance: 0.99 -> 0.78 | Helped |
| add Lproj to Lfar (ablation) | GaussianUDF · DTU | Chamfer Distance: 0.99 -> 0.88 | Helped |
| add Lproj, Lnear, and warp loss to Lfar (ablation) | GaussianUDF · DTU | Chamfer Distance: 0.99 -> 0.68 | Helped |
| add warp loss (ablation) | GaussianUDF · DTU | Chamfer Distance: 0.74 -> 0.68 | Helped |
| add Lnear self-supervision (ablation) | GaussianUDF · DTU | Chamfer Distance: 0.77 -> 0.68 | Helped |
| add Lproj (ablation) | GaussianUDF · DTU | Chamfer Distance: 0.76 -> 0.68 | Helped |
Surprising findings
- Learning UDFs directly from Gaussian centers (Only Far) degrades DTU CD to 0.99, showing noisy Gaussian point clouds are insufficient supervision.
Limitations
- Underperforms SDF-based methods on textureless structures due to UDF flexibility.
- Surface extraction from UDFs remains challenging, limiting mesh quality.
- Lacks fine detail on complex structures.
Experimental setup
Datasets: DeepFashion3D DTU NeUDF self-captured real scans
Benchmarks: DeepFashion3D DTU NeUDF
Metrics: Chamfer Distance training time
Baselines: NeuS 2DGS GOF NeuralUDF 2S-UDF VRPrior GSPull
Builds on: 2D Gaussian Splatting Neural-Pull NeuralUDF 3D Gaussian Splatting MeshUDF
Closest papers by meaning
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.93)
- GSDF: 3DGS Meets SDF for Improved Neural Rendering and Reconstruction (NeurIPS 2024 · similarity 0.90)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.89)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.89)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.89)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (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.