Research map / 3D vision / Gaussian splatting
SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View 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 SurfelSplat, a feed-forward network that regresses pixel-aligned 2D Gaussian surfels from sparse views for surface reconstruction, using Nyquist-sampling-guided low-pass surfel adaptation and cross-view feature aggregation.
Problem
Optimization-based Gaussian surfel surface reconstruction needs dense views and long per-scene optimization, while existing feed-forward pixel-aligned Gaussian methods produce geometrically inaccurate surfel orientations from sparse views.
Approach
Argues that pixel-aligned surfels exceed the Nyquist sampling rate, then applies a sampling-rate-guided Gaussian low-pass filter to adapt surfel geometry and aggregates features across views by projecting the filtered surfels into all input views with cross-attention refinement.
Key results
- Mean DTU 2-view Chamfer distance of 1.12, best among compared methods in the table (FatesGS 1.18, UFORecon 1.91, 2DGS 3.02).
- Inference in about 1 second per scene pair, reported as a 100x speedup over per-scene optimization methods.
- Nyquist analysis shows nearly all surfels exceed the threshold before adaptation and all fall within it after adaptation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Nyquist-guided surfel adaptation (ablation) | SurfelSplat · DTU evaluation scenes, 2 input views | mean Chamfer Distance | Helped |
| add cross-view feature aggregation (ablation) | SurfelSplat · DTU evaluation scenes, 2 input views | mean Chamfer Distance | Helped |
Surprising findings
- Standard feed-forward pixel-aligned surfels tend to orient parallel to the image plane, and each covers only about one pixel, so supervision is insufficient for covariance learning.
Limitations
- Pixel-aligned Gaussians scale with input resolution, yielding over 1 million Gaussians per view at 1024x1024.
- Unseen scene regions limit reconstruction; generalization is weaker than large pretrained models like VGGT or DUSt3R given limited training data.
- Predicted depth is unreliable when input views have small overlap.
Experimental setup
Datasets: RealEstate10K DTU
Benchmarks: DTU BlendedMVS
Metrics: Chamfer distance (D2S) normal rendering error inference time
Baselines: NeuS NeuSurf VolRecon UFORecon 2DGS GausSurf FatesGS
Builds on: pixelSplat 2D Gaussian Splatting Mip-Splatting MVSplat
Closest papers by meaning
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · similarity 0.90)
- Sparse2DGS: Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views (CVPR 2025 · similarity 0.90)
- Geometry Field Splatting with Gaussian Surfels (CVPR 2025 · similarity 0.90)
- HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction (ICLR 2025 · similarity 0.89)
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.88)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · 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.