Research map / 3D vision / Gaussian splatting
FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency
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 FatesGS, a 2D Gaussian splatting framework for sparse-view surface reconstruction that uses patch-based monocular depth ranking, depth smoothing, and multi-level multi-view feature projection losses, trained in minutes without pre-training.
Problem
Surface reconstruction from sparse input views with Gaussian splatting, where primitives overfit few training views, producing floaters and incomplete surfaces; existing sparse methods need days of pre-training or hours per scene.
Approach
Builds on 2D Gaussian splatting and adds intra-view depth consistency (patch-wise ranking from monocular depth plus a depth smoothness loss) for coarse geometry, and multi-view feature alignment of reprojected depth points using frozen Vis-MVSNet features at multiple scales with a visibility mask for fine detail.
Key results
- Achieves mean Chamfer Distance 0.92 on DTU 3-view large-overlap setting, the best among compared methods (UFORecon 0.99, NeuSurf 0.99, 2DGS 1.69).
- Per-scene training takes about 14 minutes on a single RTX 3090 with 4 GB GPU memory, versus 6 to 14 hours for MonoSDF and NeuSurf, claimed as a 60x to 200x speedup.
- Depth evaluation on DTU: 77.35% of pixels within 1mm for seen views versus 35.74% for 2DGS and 5.01% for Marigold.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FatesGS instead of UFORecon | DTU large-overlap surface reconstruction · DTU large-overlap (15-scene mean) | Chamfer Distance: 0.99 -> 0.92 | Helped |
| add multi-view feature loss to ranking and smoothing losses (ablation) | FatesGS · DTU little-overlap | Mean Chamfer Distance: 2.56 -> 1.37 | Helped |
| add smoothing loss to ranking and feature losses (ablation) | FatesGS · DTU little-overlap | Mean Chamfer Distance: 1.56 -> 1.37 | Helped |
| add ranking loss to smoothing and feature losses (ablation) | FatesGS · DTU little-overlap | Mean Chamfer Distance: 1.62 -> 1.37 | Helped |
| add ranking and smoothing losses without feature loss (ablation) | FatesGS · DTU little-overlap | Mean Chamfer Distance: 2.47 -> 2.56 | Hurt |
| increase training views from 3 to 6 (ablation) | FatesGS · DTU | Mean Chamfer Distance: 0.92 -> 0.85 | Helped |
| increase training views from 6 to 9 (ablation) | FatesGS · DTU | Mean Chamfer Distance: 0.85 -> 0.79 | Helped |
| increase training views from 9 to full views (ablation) | FatesGS · DTU | Mean Chamfer Distance: 0.79 -> 0.61 | Helped |
| use FatesGS rather than Marigold for depth prediction | DTU depth prediction, seen and unseen views · DTU depth prediction | Abs.: 15.58 / 15.12 -> 2.41 / 3.43 | Helped |
Surprising findings
- Using only the ranking and smoothing losses performs worse than the baseline with none of the three losses (CD 2.56 vs 2.47 on little-overlap setting).
Limitations
- Depends on monocular depth (Marigold) and a frozen feature extractor (Vis-MVSNet) as external priors.
- Relies on COLMAP initialization and known camera poses.
Experimental setup
Datasets: DTU BlendedMVS
Benchmarks: DTU BlendedMVS
Metrics: Chamfer Distance depth threshold percentage absolute error relative error training time GPU memory
Baselines: COLMAP TransMVSNet SparseNeuS VolRecon ReTR C2F2NeuS GenS UFORecon NeuS VolSDF MonoSDF NeuSurf
Builds on: 2D Gaussian Splatting 3D Gaussian Splatting Marigold Vis-MVSNet
Closest papers by meaning
- Sparse2DGS: Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views (CVPR 2025 · similarity 0.93)
- SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction (NeurIPS 2025 · similarity 0.90)
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.90)
- TranSplat: Generalizable 3D Gaussian Splatting from Sparse Multi-View Images with Transformers (AAAI 2025 · similarity 0.90)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.90)
- DepthSplat: Connecting Gaussian Splatting and Depth (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.