Research map / 3D vision / Face and shape
SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration
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 SuperNormal, a fast multi-view normal-integration method that optimizes a hash-encoded neural SDF with directional finite differences and patch-based ray marching to reconstruct detailed surfaces from photometric-stereo normal maps.
Problem
Multi-view photometric stereo (MVPS) normal fusion is slow (hours per object) and existing methods smooth out high-frequency surface details from normal maps.
Approach
Supervise a multi-resolution hash-encoded neural SDF with rendered normals and masks, computing SDF gradients via directional finite difference whose samples are reused from volume rendering through patch-based ray marching, avoiding double backpropagation and extra SDF queries.
Key results
- On DiLiGenT-MV, SuperNormal achieves average L2 Chamfer distance 0.194 mm and F-score 0.962 at 48 seconds per object, the best geometry accuracy among compared MVPS methods.
- Directional finite difference total runtime 48.3 s vs 78.1 s for automatic differentiation and 124.6 s for finite difference, with no significant accuracy loss.
- Applying DFD and patch-based sampling to Neuralangelo reduced training time by 25% on a DTU scene.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-resolution hash encoding instead of positional encoding (ablation) | SuperNormal · DiLiGenT-MV (5-object average), SDM-UniPS normals | L2 Chamfer distance [mm]: 0.228 -> 0.194 | Helped |
| use multi-resolution hash encoding instead of positional encoding (ablation) | SuperNormal · DiLiGenT-MV (5-object average), SDM-UniPS normals | F-score (τF = 0.5 mm): 0.942 -> 0.962 | Helped |
| use SDM-UniPS normal maps instead of SDPS normal maps (ablation) | SuperNormal (hash encoding) · DiLiGenT-MV (5-object average) | L2 Chamfer distance [mm]: 0.281 -> 0.194 | Helped |
| use SDM-UniPS normal maps instead of SDPS normal maps (ablation) | SuperNormal (hash encoding) · DiLiGenT-MV (5-object average) | F-score (τF = 0.5 mm): 0.860 -> 0.962 | Helped |
| use GT normal maps instead of SDM-UniPS normal maps (ablation) | SuperNormal (hash encoding) · DiLiGenT-MV (5-object average) | L2 Chamfer distance [mm]: 0.194 -> 0.114 | Helped |
| replace directional finite difference with finite difference (ablation) | SuperNormal · DiLiGenT-MV (5-object average) | Total runtime: 48.3 -> 124.6 | Hurt |
| replace directional finite difference with automatic differentiation (ablation) | SuperNormal · DiLiGenT-MV (5-object average) | Total runtime: 48.3 -> 78.1 | Hurt |
| replace directional finite difference with finite difference (ablation) | SuperNormal · DiLiGenT-MV (5-object average) | CD: 0.194 -> 0.192 | No clear effect |
| replace directional finite difference with automatic differentiation (ablation) | SuperNormal · DiLiGenT-MV (5-object average) | F-Score: 0.961 -> 0.961 | No clear effect |
| use SuperNormal instead of B-MVPS | DiLiGenT-MV · DiLiGenT-MV (5-object average) | L2 Chamfer distance [mm]: 0.203 -> 0.194 | Helped |
| use SuperNormal instead of MVPSNet | DiLiGenT-MV · DiLiGenT-MV (5-object average) | F-score (τF = 0.5 mm): 0.865 -> 0.962 | Helped |
| use SuperNormal instead of MVPSNet | DiLiGenT-MV · DiLiGenT-MV (5-object average) | Runtime: 37 secs -> 48 secs | Hurt |
Surprising findings
- Method is sensitive to inaccurate normal maps from monocular estimators such as OmniData, failing to reach good quality without photometric stereo normals.
Limitations
- Produces fault-like artifacts regardless of gradient computation method
- Depends on mask supervision and high-quality photometric stereo normal maps
- Performs poorly with monocular normal estimation such as OmniData
Experimental setup
Datasets: DiLiGenT-MV DTU
Benchmarks: DiLiGenT-MV
Metrics: L2 Chamfer distance F-score runtime
Baselines: R-MVPS B-MVPS MVPSNet UA-MVPS PS-NeRF MVAS NeuS2 Neuralangelo
Builds on: NeuS Instant-NGP Neuralangelo SDM-UniPS NerfAcc
Closest papers by meaning
- NeAT: Learning Neural Implicit Surfaces With Arbitrary Topologies From Multi-View Images (CVPR 2023 · similarity 0.89)
- MNSRNet: Multimodal Transformer Network for 3D Surface Super-Resolution (CVPR 2022 · similarity 0.89)
- PMNI: Pose-free Multi-view Normal Integration for Reflective and Textureless Surface Reconstruction (CVPR 2025 · similarity 0.88)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.87)
- RNb-NeuS: Reflectance and Normal-based Multi-View 3D Reconstruction (CVPR 2024 · similarity 0.87)
- SC-NeuS: Consistent Neural Surface Reconstruction from Sparse and Noisy Views (AAAI 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.