Research map / Point clouds / Shape and surface completion
CMG-Net: Robust Normal Estimation for Point Clouds via Chamfer Normal Distance and Multi-Scale Geometry
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 Chamfer Normal Distance (CND), a loss/metric that replaces annotated normals with the normal of the closest clean point, and CMG-Net, a multi-scale graph/hierarchical network for robust point cloud normal estimation.
Problem
Learning-based point cloud normal estimation degrades on noisy point clouds because annotated ground-truth normals become inconsistent with the noisy local neighborhood geometry, and fixed scale choices hurt robustness and detail preservation.
Approach
Replace the fixed annotated normal in the loss/metric with the normal of the nearest point on the noise-free surface (CND), and combine multi-scale kNN local feature aggregation (with attentional fusion), hierarchical local-global fusion, and position-aware weighted normal prediction in a network.
Key results
- On PCPNet, the method achieves average CND of 8.35 versus 8.59 for SHS-Net, with the best results under noise (e.g., 8.13 at sigma=0.12%).
- On the real-world SceneNN dataset, average CND is 8.87 versus 9.25 for SHS-Net.
- Applying the CND-modified loss to DeepFit, HSurf-Net and PCPNet reduces their average RMSE (e.g., DeepFit 10.62 to 10.48 average, PCPNet 13.52 to 12.19).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CMG-Net instead of SHS-Net | PCPNet · PCPNet | RMSE (Ave.): 9.96 -> 9.93 | Helped |
| Use CMG-Net instead of SHS-Net | PCPNet · PCPNet | CND (Ave.): 8.59 -> 8.35 | Helped |
| Use CMG-Net instead of HSurf-Net | SceneNN · SceneNN | CND (Ave.): 9.02 -> 8.87 | Helped |
| Add CND-modified loss (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.41 -> 8.35 | Helped |
| Add Local Feature Extraction (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.52 -> 8.35 | Helped |
| Use multi-scale rather than single-scale Local Feature Extraction (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.50 -> 8.35 | Helped |
| Add Attentional Feature Fusion (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.44 -> 8.35 | Helped |
| Add Hierarchical Architecture (ablation) | CMG-Net · PCPNet | CND (Ave.): 9.07 -> 8.35 | Helped |
| Add Multi-scale Global Feature (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.39 -> 8.35 | Helped |
| Add Local Feature to hierarchical architecture (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.49 -> 8.35 | Helped |
| Add Position Feature Fusion (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.42 -> 8.35 | Helped |
| Add Weighted Normal Prediction (ablation) | CMG-Net · PCPNet | CND (Ave.): 8.54 -> 8.35 | Helped |
Surprising findings
- Denoising pre-processing before normal estimation worsened accuracy (average 15.99 vs 12.30 without it).
Limitations
- Not real-time capable (about 4 s per 10K-point scene).
- Still depends on annotated training data, like prior approaches.
- Gains over strongest baselines on clean data are small; the largest gains are under noise.
Experimental setup
Datasets: PCPNet SceneNN
Benchmarks: PCPNet SceneNN
Metrics: RMSE CND AUC PGP MSAE SMD CD
Baselines: PCA n-jet PCPNet Nesti-Net DeepFit AdaFit GraphFit HSurf-Net Du et al. SHS-Net
Builds on: PCPNet DeepFit HSurf-Net QSTN AdaFit
Closest papers by meaning
- High-quality Point Cloud Oriented Normal Estimation via Hybrid Angular and Euclidean Distance Encoding (CVPR 2025 · similarity 0.87)
- An Explicit Frame Construction for Normalizing 3D Point Clouds (ICML 2024 · similarity 0.86)
- Consistent Normal Orientation for 3D Point Clouds via Least Squares on Delaunay Graph (CVPR 2025 · similarity 0.85)
- Dual Manifold Regularization Steered Robust Representation Learning for Point Cloud Analysis (AAAI 2025 · similarity 0.85)
- LiSu: A Dataset and Method for LiDAR Surface Normal Estimation (CVPR 2025 · similarity 0.85)
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · 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.