Research map / Point clouds / Shape and surface completion
High-quality Point Cloud Oriented Normal Estimation via Hybrid Angular and Euclidean Distance Encoding
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 HAE, a three-stage LiDAR-aware network that estimates and orients point cloud normals by fusing Euclidean coordinates with ray (angular) features, plus a virtual-scanner dataset with ground-truth normals.
Problem
Estimating accurate oriented normals from LiDAR point clouds, which are noisy and unevenly distributed in Euclidean space, while existing methods rely only on local xyz geometry and cannot exploit scanner ray information.
Approach
Combine Euclidean and angular (ray direction) spaces: an angular distance encoding module refines unoriented normals using ray directions and coordinates, and a ray feature fusion module classifies whether each normal should be flipped using ray-normal dot products, trained on virtually scanned data with ground-truth normals.
Key results
- On FamousShape, oriented-normal average modified RMSE of 20.23 vs 29.27 (GraphFit) and 29.41 (CMG-Net), a large reduction.
- On ScanNet, oriented-normal average modified RMSE of 17.32 vs 18.27 for CMG-Net (best baseline).
- Ablation on FamousShape shows removing the ray feature fusion module raises oriented average RMSE from 20.23 to 25.35.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HAE instead of CMG-Net | oriented normal estimation · FamousShape | modified RMSE (Average): 29.41 -> 20.23 | Helped |
| use HAE instead of CMG-Net | oriented normal estimation · ScanNet | modified RMSE (Average): 18.27 -> 17.32 | Helped |
| add angular distance encoding (ablation) | HAE · FamousShape | modified RMSE (oriented normals, Average): 20.59 -> 20.23 | Helped |
| add ray feature fusion (ablation) | HAE · FamousShape | modified RMSE (oriented normals, Average): 25.35 -> 20.23 | Helped |
| add Lsin loss (ablation) | HAE · FamousShape | modified RMSE (oriented normals, Average): 20.30 -> 20.23 | Helped |
| add Lmse loss (ablation) | HAE · FamousShape | modified RMSE (oriented normals, Average): 20.45 -> 20.23 | Helped |
Surprising findings
- SHS-Net gives very large oriented-normal errors on ScanNet (about 39 RMSE) despite competitive unoriented results, indicating orientation failures.
Limitations
- Not evaluated on point clouds not directly acquired by sensors, such as MVS-generated clouds.
- Orientation relies on rays only illuminating object exteriors; noise can still penetrate thin structures.
- Datasets are simulated from meshes; real-scan evaluation is only qualitative on one chemical plant scene.
Experimental setup
Datasets: FamousShape ScanNet v2
Benchmarks: FamousShape ScanNet
Metrics: modified RMSE (unoriented) modified RMSE (oriented)
Baselines: PCA N-jet GraphFit Du et al. SHS-Net CMG-Net
Builds on: PCPNet DGCNN (EdgeConv) SHS-Net PCA normal estimation
Closest papers by meaning
- LiSu: A Dataset and Method for LiDAR Surface Normal Estimation (CVPR 2025 · similarity 0.87)
- 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 0.87)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.86)
- An Explicit Frame Construction for Normalizing 3D Point Clouds (ICML 2024 · similarity 0.86)
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.85)
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.