Research map / Point clouds / Shape and surface completion
Refine-Net: Normal Refinement Neural Network for Noisy Point Clouds
Haoran Zhou, Honghua Chen, Yingkui Zhang, Mingqiang Wei, Haoran Xie, Jun Wang, Tong Lü, Jing Qin, Xiao-Ping Steven Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Refine-Net refinement on MFPS initial normals | MFPS · Wang et al. synthetic | mean angular error: 4.75 -> 3.87 | Improved |
| add Refine-Net refinement on Nesti-Net normals | Nesti-Net · Wang et al. synthetic | mean angular error: 4.55 -> 4.11 | Improved |
| add Refine-Net refinement on PCPNet-ms normals | PCPNet-ms · Wang et al. synthetic | mean angular error: 7.06 -> 5.64 | Improved |
| add fitting patch selection to MFPS (ablation) | MFPS · Wang et al. synthetic | mean angular error: 5.76 -> 4.75 | Improved |
| add point and height-map feature modules to normals (ablation) | Refine-Net · Wang et al. synthetic | mean angular error: 4.40 -> 3.87 | Improved |
| add point module to normals&HMPs (ablation) | Refine-Net · Wang et al. synthetic | mean angular error: 4.31 -> 3.87 | Improved |
| add height-map module to normals&points (ablation) | Refine-Net · Wang et al. synthetic | mean angular error: 3.97 -> 3.87 | Improved |
| use weight matrix connection instead of concatenation (ablation) | Refine-Net · Wang et al. synthetic | mean angular error: 4.04 -> 3.87 | Improved |
| use weight matrix connection instead of residual (ablation) | Refine-Net · Wang et al. synthetic | mean angular error: 4.03 -> 3.87 | Improved |
| use weight matrix instead of translation matrix (ablation) | Refine-Net · Wang et al. synthetic | mean angular error: 3.95 -> 3.87 | Improved |
| use L2 instead of L1 loss (ablation) | Refine-Net · PCPNet dataset (noise 0.012) | angular RMS error: 23.50 -> 22.57 | Improved |
| use L2 instead of L1 loss (ablation) | Refine-Net · PCPNet dataset (noise none) | angular RMS error: 6.05 -> 6.27 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Nesti-Net — Nesti-Net: Normal Estimation for Unstructured 3D Point Clouds Using Convolutional Neural Networks (CVPR 2019) · better on 8 of 9 shared results
- DeepFit — DeepFit: 3D Surface Fitting via Neural Network Weighted Least Squares (ECCV 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Geometry and Learning Co-Supported Normal Estimation for Unstructured Point Cloud (CVPR 2020 · similarity 0.90)
- Learning Normals of Noisy Points by Local Gradient-Aware Surface Filtering (ICCV 2025 · similarity 0.90)
- HSurf-Net: Normal Estimation for 3D Point Clouds by Learning Hyper Surfaces (NeurIPS 2022 · similarity 0.90)
- NeuralGF: Unsupervised Point Normal Estimation by Learning Neural Gradient Function (NeurIPS 2023 · similarity 0.90)
- CMG-Net: Robust Normal Estimation for Point Clouds via Chamfer Normal Distance and Multi-Scale Geometry (AAAI 2024 · similarity 0.87)
- Consistent Normal Orientation for 3D Point Clouds via Least Squares on Delaunay Graph (CVPR 2025 · similarity 0.86)
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.
