Research map / 3D vision / Face and shape
NeuSurf: On-Surface Priors for Neural Surface Reconstruction from Sparse Input 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.
Proposes NeuSurf, a sparse-view neural SDF reconstruction method that uses SfM on-surface points as training-free priors, combining a UDF-based global geometry alignment with a feature-consistency local refinement loss.
Problem
Neural implicit surface reconstruction from sparse input views produces collapsed, incomplete or noisy geometry, and existing generalizable priors need costly large-scale training and only work for a fixed sparse-view setting.
Approach
Derive priors directly from SfM on-surface points without extra training data: fit a UDF network to these points via a CAP-UDF-style moving operation as a coarse global constraint, and add a local loss enforcing consistency of projected features of on-surface points (including differentiable ray-surface pseudo points) across seen and unseen views.
Key results
- On DTU little-overlap setting, NeuNSurf achieves mean Chamfer Distance 1.35 versus 3.02 for VolRecon and 1.86 for MonoSDF.
- On DTU large-overlap setting, mean Chamfer Distance 0.99 versus 1.38 for VolRecon and 1.52 for COLMAP.
- Ablation on little-overlap DTU: removing both losses gives mean CD 2.46, L_global alone 1.67, L_local alone 1.96, both 1.35.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeuSurf instead of MonoSDF | DTU little-overlap reconstruction · DTU little-overlap | Chamfer Distance: 1.86 -> 1.35 | Helped |
| Use NeuSurf instead of VolRecon | DTU large-overlap reconstruction · DTU large-overlap | Chamfer Distance: 1.38 -> 0.99 | Helped |
| Add global and local geometry losses (ablation) | NeuSurf · DTU little-overlap | Mean Chamfer Distance: 2.46 -> 1.35 | Helped |
| Add global geometry loss (ablation) | NeuSurf with local geometry loss · DTU little-overlap | Mean Chamfer Distance: 1.96 -> 1.35 | Helped |
| Add local geometry loss (ablation) | NeuSurf with global geometry loss · DTU little-overlap | Mean Chamfer Distance: 1.67 -> 1.35 | Helped |
Limitations
- Relies on COLMAP SfM point clouds, which can be fragmented or sparse with sparse views
- Per-scene optimization with 300k iterations is expensive
Experimental setup
Datasets: DTU BlendedMVS
Benchmarks: DTU BlendedMVS
Metrics: Chamfer Distance
Baselines: COLMAP SparseNeuS VolRecon NeuS VolSDF MonoSDF
Builds on: NeuS CAP-UDF MonoSDF SparseNeuS
Closest papers by meaning
- SC-NeuS: Consistent Neural Surface Reconstruction from Sparse and Noisy Views (AAAI 2024 · similarity 0.92)
- NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images (CVPR 2024 · similarity 0.88)
- NeuralIndicator: Implicit Surface Reconstruction from Neural Indicator Priors (ICML 2024 · similarity 0.88)
- Sparis: Neural Implicit Surface Reconstruction of Indoor Scenes from Sparse Views (AAAI 2025 · similarity 0.87)
- UFORecon: Generalizable Sparse-View Surface Reconstruction from Arbitrary and Unfavorable Sets (CVPR 2024 · similarity 0.87)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.87)
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.