Research map / 3D vision / Gaussian splatting
Sparse2DGS: Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views
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.
Sparse2DGS initializes 2D Gaussian Splatting with dense MVS (CLMVSNet) points and adds geometry-prioritized losses (fixed-color MVS feature splatting, reparameterized disk-sampling cross-view regularization, and NCC-based selective position updates) for sparse-view surface reconstruction.
Problem
Surface reconstruction from extremely sparse input views, where SfM initialization is too sparse and plain combination of MVS points with Gaussian Splatting yields suboptimal geometry due to ill-posed optimization.
Approach
Use unsupervised MVS (CLMVSNet) for dense Gaussian initialization, fix per-Gaussian color and supervise geometry via splatted MVS features, regularize disk primitives through reparameterized point sampling with cross-view feature consistency, and selectively replace Gaussian positions with rendered depth when patch NCC improves instead of adaptive density control.
Key results
- Achieves Chamfer Distance 1.13 on DTU with 3 views, versus 2DGS 2.81, GOF 2.82, PGSR 2.08, and SparseNeuS 1.27.
- Plain 2DGS+CLMVSNet combination gives CD 1.27, not surpassing CLMVSNet alone (1.26), while the full method reaches 1.13.
- Training takes about 10 minutes versus 25 minutes for SparseNeuS, roughly 2x faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace sparse SfM initialization with MVS point cloud initialization (ablation) | 2DGS · DTU 3-view | Chamfer Distance: 2.813 -> 1.273 | Helped |
| Fix color and use MVS feature splatting for geometric supervision (ablation) | Sparse2DGS · DTU | Average: 1.27 -> 1.21 | Helped |
| Add Direct Gaussian Primitive Regularization (ablation) | Sparse2DGS · DTU | Chamfer Distance: 1.201 -> 1.142 | Helped |
| Replace Adaptive Density Control with Selective Gaussian Update (ablation) | 2DGS · DTU | Average: 1.234 -> 1.152 | Helped |
| Use Adaptive Density Control instead of the baseline update strategy (ablation) | 2DGS · DTU | Average: 1.201 -> 1.234 | Hurt |
| Increase disk-sampling points from 9 to 81 (ablation) | Sparse2DGS · DTU | Average: 1.145 -> 1.139 | Helped |
| Use Sparse2DGS instead of NeuSurf | DTU 3-view reconstruction · DTU 3-view (mean) | Chamfer Distance: 0.99 -> 1.13 | Hurt |
| Use Sparse2DGS instead of 2DGS | DTU 3-view reconstruction · DTU 3-view (mean) | Chamfer Distance: 2.81 -> 1.13 | Helped |
| Use Sparse2DGS instead of SparseNeus | DTU 3-view reconstruction · DTU 3-view | training time: 25mins -> 10mins | Helped |
Surprising findings
- Plainly combining MVS point clouds with 2DGS does not beat CLMVSNet alone (CD 1.27 vs 1.26).
- Adaptive density control from 2DGS degrades performance under sparse views (CD 1.23 vs baseline 1.20).
Limitations
- MVS degradation under occlusions propagates into the reconstruction.
- Evaluated mainly on DTU with 3 views; qualitative only on BlendedMVS.
Experimental setup
Datasets: DTU BlendedMVS
Benchmarks: DTU BlendedMVS
Metrics: Chamfer Distance training time
Baselines: 2DGS GOF PGSR SparseNeuS VolRecon ReTR NeuSurf CLMVSNet Colmap NeuS VolSDF C2F2Neus
Builds on: 2D Gaussian Splatting CLMVSNet 3D Gaussian Splatting SparseNeuS
Closest papers by meaning
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · similarity 0.93)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.91)
- TranSplat: Generalizable 3D Gaussian Splatting from Sparse Multi-View Images with Transformers (AAAI 2025 · similarity 0.91)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.90)
- SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction (NeurIPS 2025 · similarity 0.90)
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.90)
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.