Research map / 3D vision / NeRF and radiance fields
H2O-SDF: Two-phase Learning for 3D Indoor Reconstruction using Object Surface Fields
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 H2O-SDF, a two-phase NeuS-based method for indoor 3D reconstruction that uses a normal-uncertainty loss re-weighting and an Object Surface Field (OSF) with OSF-guided sampling to capture object details and mitigate vanishing gradients.
Problem
Indoor 3D scene reconstruction from multi-view images struggles to jointly capture smooth low-frequency room layouts and high-frequency small object surfaces; SDF-based methods suffer vanishing gradients and smoothness bias on thin structures.
Approach
Two-phase learning: holistic surface learning re-weights normal and color losses by monocular normal uncertainty, then an Object Surface Field (a per-point object-surface probability) is trained with 2D mask, 3D object surface and refinement losses that mutually guide the SDF, plus OSF-weighted importance sampling.
Key results
- On ScanNet, H2O-SDF achieves F-score 0.799 vs. 0.755 for HelixSurf (previous best) and Acc 0.032 / Comp 0.0373, the best on all geometry metrics.
- On Replica, F-score 0.957 vs. 0.934 for MonoSDF; on 7-Scenes, F-score 0.447 vs. 0.410 for NeuRIS.
- Ablation on 12 ScanNet scenes: F-score rises from 0.353 (NeuS) to 0.766 with normal priors, to 0.800 with the full model including OSF and OGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add normal priors (ablation) | NeuS · ScanNet (12 scenes) | F-score: 0.353 -> 0.766 | Helped |
| add normal-uncertainty reweighting (ablation) | NeuS with normal priors · ScanNet (12 scenes) | F-score: 0.766 -> 0.776 | Helped |
| add 2D object surface loss (ablation) | NeuS with normal priors and reweighting · ScanNet (12 scenes) | F-score: 0.776 -> 0.779 | Helped |
| add 3D object surface loss (ablation) | NeuS with normal priors, reweighting, and 2D OSF loss · ScanNet (12 scenes) | F-score: 0.779 -> 0.784 | Helped |
| add OSF refinement loss (ablation) | NeuS with normal priors, reweighting, 2D OSF loss, and 3D OSF loss · ScanNet (12 scenes) | F-score: 0.784 -> 0.787 | Helped |
| add OSF-guided sampling (ablation) | H2O-SDF with normal priors, reweighting, and all OSF losses · ScanNet (12 scenes) | F-score: 0.787 -> 0.800 | Helped |
Surprising findings
- Adding normal priors to NeuS alone raises F-score from 0.353 to 0.766, so most of the gain comes from the normal prior before the OSF components.
Limitations
- Performance on some scenes (scene0603, scene0616) is slightly worse than NeuRIS or MonoSDF, attributed partly to missing ground-truth mesh areas.
- Relies on pretrained normal, segmentation and MVS priors whose errors can propagate into the SDF.
- Training uses the slower NeuS base; faster radiance-field backbones are not yet integrated.
Experimental setup
Datasets: ScanNet Replica 7-Scenes
Benchmarks: ScanNet Replica 7-Scenes
Metrics: Accuracy Completeness Precision Recall F-score normal mean normal median normal RMSE
Baselines: COLMAP ACMP NeRF VolSDF NeuS ManhattanSDF NeuRIS MonoSDF HelixSurf
Builds on: NeuS MonoSDF NeuRIS ObjectSDF I2-SDF
Closest papers by meaning
- NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation (CVPR 2024 · similarity 0.89)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.88)
- ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction (ICLR 2025 · similarity 0.87)
- Latent Radiance Fields with 3D-aware 2D Representations (ICLR 2025 · similarity 0.85)
- S-INF: Towards Realistic Indoor Scene Synthesis via Scene Implicit Neural Field (AAAI 2025 · similarity 0.85)
- 2S-UDF: A Novel Two-stage UDF Learning Method for Robust Non-watertight Model Reconstruction from Multi-view Images (CVPR 2024 · similarity 0.84)
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.