Research map / 3D vision / Face and shape
NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation
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 NC-SDF, a neural SDF indoor reconstruction method that learns view-dependent rotation corrections for monocular normal priors, plus informative (Canny-guided) pixel sampling and a hybrid MLP-voxel geometry model.
Problem
Indoor neural implicit surface reconstruction relying on monocular normal priors suffers from view-dependent, multi-view-inconsistent biases in those priors, leading to noisy or missing surfaces and lost details in texture-less and detailed regions.
Approach
Model view-dependent biases in monocular normal priors as predicted rotation angles (about x, y, z) applied to the SDF normal, learned jointly with radiance and SDF fields in a second training stage; combine with Canny-based informative pixel sampling and an MLP+voxel-grid hybrid geometry model.
Key results
- On ScanNet, NC-SDF reaches F-score 0.781 vs. 0.756 for HelixSurf and 0.748 for MonoSDF (MLP).
- On ICL-NUIM, F-score 0.852 vs. 0.650 for MonoSDF (MLP) and 0.626 for NeuRIS.
- Ablation: hybrid+IPS improves F-score by 2.18%; adding normal compensation adds a further 4.37% (ScanNet).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hybrid geometry modeling instead of MLP-only geometry (ablation) | NC-SDF · ScanNet | F-score: 0.733 -> 0.745 | Helped |
| Add informative pixel sampling to MLP geometry (ablation) | NC-SDF · ScanNet | F-score: 0.733 -> 0.742 | Helped |
| Add informative pixel sampling to hybrid geometry (ablation) | NC-SDF · ScanNet | F-score: 0.745 -> 0.749 | Helped |
| Add view-dependent normal compensation (ablation) | NC-SDF with hybrid geometry and informative pixel sampling · ScanNet | F-score: 0.749 -> 0.781; 4.37% | Helped |
Limitations
- Performance depends on quality of monocular normal priors from pretrained network
- Per-scene optimization with long training (80k+ iterations)
- Evaluated on only four scenes per dataset
Experimental setup
Datasets: ScanNet ICL-NUIM
Benchmarks: ScanNet ICL-NUIM
Metrics: accuracy completeness precision recall F-score
Baselines: COLMAP VolSDF NeuS ManhattanSDF HelixSurf NeuRIS MonoSDF (MLP) MonoSDF (Grid)
Builds on: NeuS MonoSDF VolSDF NeRF
Closest papers by meaning
- ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction (ICLR 2025 · similarity 0.90)
- H2O-SDF: Two-phase Learning for 3D Indoor Reconstruction using Object Surface Fields (ICLR 2024 · similarity 0.89)
- 3D Reconstruction with Generalizable Neural Fields using Scene Priors (ICLR 2024 · similarity 0.88)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.88)
- Learning Neural Exposure Fields for View Synthesis (NeurIPS 2025 · similarity 0.87)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (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.