Research map / 3D vision / Depth and stereo
SC-NeuS: Consistent Neural Surface Reconstruction from Sparse and Noisy 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.
Supplementary material for SC-NeuS, a neural surface reconstruction method for sparse, noisy-pose views, covering a refined DTU Chamfer evaluation (ICP after Sim3), COLMAP comparison, Geo-NeuS-BARF comparison, and timing.
Problem
Neural surface reconstruction via volume rendering needs dense, accurately posed views; the supplement addresses fair evaluation of reconstruction quality and extra comparisons for sparse views with noisy camera poses.
Approach
Refines the standard DTU Chamfer Distance evaluation by applying ICP after the Sim3 alignment so CD reflects surface quality rather than camera-pose estimation accuracy, and jointly optimizes neural surface and camera poses with a view-consistent patch-warping (NCC) loss.
Key results
- On DTU with noisy poses, the method reaches mean CD 1.41 versus 3.45 for the w/o L_ncc variant and 4.46 for COLMAP.
- Mean camera translation RMSE of 0.15 and rotation RMSE of 0.12 degrees with noisy initialization, versus 0.22 and 0.29 for COLMAP.
- Geo-NeuS with noisy initialization reaches CD 5.00 and translation RMSE 8.60, far worse than the proposed method (CD 1.41, RMSE 0.15).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the view-consistent patch-warping loss L_ncc to the reprojection-loss system (ablation) | SC-NeuS · DTU | translation error: 0.23 -> 0.15 | Helped |
| Add the view-consistent patch-warping loss L_ncc to the reprojection-loss system (ablation) | SC-NeuS · DTU | rotation error (°): 0.30 -> 0.12 | Helped |
| Add the view-consistent patch-warping loss L_ncc to the reprojection-loss system (ablation) | SC-NeuS · DTU | Chamfer Distance: 3.45 -> 1.41 | Helped |
| Use SC-NeuS rather than COLMAP | DTU sparse-view reconstruction · DTU | Chamfer Distance: 4.46 -> 1.41 | Helped |
| Use SC-NeuS rather than Geo-NeuS with noisy initialization | DTU reconstruction with noisy initialization · DTU, noisy initialization | Chamfer Distance: 5.00 -> 1.41 | Helped |
Surprising findings
- The standard Sim3-based DTU Chamfer evaluation can unfairly penalize reconstructions whose poses are poorly estimated, motivating an extra ICP refinement step.
Limitations
- Reconstruction quality depends on 2D feature-match quality, similar to COLMAP.
- ICP refinement can fail without good pose initialization, so the Sim3 alignment is used as fallback.
- Training takes about 4.5 hours per object on an RTX 3090.
Experimental setup
Datasets: DTU BlendedMVS
Benchmarks: DTU BlendedMVS
Metrics: Chamfer Distance camera translation RMSE camera rotation RMSE
Baselines: COLMAP BARF IDR NeuS-BARF Geo-NeuS-BARF
Builds on: NeuS BARF Geo-NeuS
Closest papers by meaning
- NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction (NeurIPS 2021 · similarity 0.93)
- NeuSurf: On-Surface Priors for Neural Surface Reconstruction from Sparse Input Views (AAAI 2024 · similarity 0.92)
- Geo-Neus: Geometry-Consistent Neural Implicit Surfaces Learning for Multi-view Reconstruction (NeurIPS 2022 · similarity 0.90)
- PMNI: Pose-free Multi-view Normal Integration for Reflective and Textureless Surface Reconstruction (CVPR 2025 · similarity 0.87)
- NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images (CVPR 2024 · similarity 0.87)
- SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration (CVPR 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.