Research map / Point clouds / Shape and surface completion
EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features
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 EdgeMovingNet, a point-cloud network jointly regressing point normal, point-to-edge direction and distance, plus a refinement and vertex-relocation pipeline for edge-preserving surface reconstruction from unoriented sparse points.
Problem
Reconstructing CAD-style meshes with sharp edges from sparse point clouds without normals, where few points lie exactly on edges, causing artifacts in reconstruction.
Approach
Jointly predict per-point normal, point-to-edge direction and point-to-edge distance with a global encoder-decoder, use a surface loss and tangent loss to enforce edge geometry, then generate and optimize edge points by neighbor midpoints and predicted normals, and relocate Poisson mesh vertices onto those edge points.
Key results
- On the ABC test set, the method reports CD 0.703 and HD 1.368 (x10^-2), the best among compared methods, with F-score 0.993 versus 0.988 for Voronoi.
- On ModelNet, F-score 0.982 versus 0.973 for Voronoi, the next best.
- Ablation shows removing the refinement process raises CD by 0.122 and HD by 0.368 (x10^-2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EdgeMovingNet with joint point-to-edge direction, distance, and normal prediction plus refinement | EdgeMovingNet · ABC | CD: 0.709 -> 0.703 | Helped |
| use EdgeMovingNet with joint point-to-edge direction, distance, and normal prediction plus refinement | EdgeMovingNet · ABC | HD: 1.401 -> 1.368 | Helped |
| use EdgeMovingNet with joint point-to-edge direction, distance, and normal prediction plus refinement | EdgeMovingNet · ABC | F-score: 0.988 -> 0.993 | Helped |
| use EdgeMovingNet reconstruction | EdgeMovingNet · ABC, 1% noise | HD: 3.311 -> 1.858 | Helped |
| use EdgeMovingNet reconstruction | EdgeMovingNet · ABC, 2% noise | HD: 4.294 -> 1.902 | Helped |
| use EdgeMovingNet reconstruction | EdgeMovingNet · ModelNet | CD: 0.965 -> 0.919 | Helped |
| use EdgeMovingNet reconstruction | EdgeMovingNet · ShapeNet | CD: 1.219 -> 1.186 | Helped |
| add explicit point-to-edge direction and distance prediction instead of regressing one displacement vector (ablation) | EdgeMovingNet · ABC | CD: +0.152 | Helped |
| add explicit point-to-edge direction and distance prediction instead of regressing one displacement vector (ablation) | EdgeMovingNet · ABC | HD: +1.384 | Helped |
| add surface loss (ablation) | EdgeMovingNet · ABC | CD: +0.017 | Helped |
| add surface loss (ablation) | EdgeMovingNet · ABC | HD: +0.047 | Helped |
| add tangent loss (ablation) | EdgeMovingNet · ABC | CD: +0.016 | Helped |
Surprising findings
- Learning-based POCO and NKSR underperform classical Voronoi and Poisson on edge-focused metrics in several settings.
Limitations
- Performance drops on very thin structures or intricate details such as keyboards, due to the non-patch-based design.
- Performance degrades with heavy noise or outliers, requiring a pre-denoising step.
- Only mildly robust to noise; the default model is not trained with noisy data.
Experimental setup
Datasets: ABC ModelNet ShapeNet AIM@SHAPE-VISIONAIR
Benchmarks: ABC ModelNet ShapeNet
Metrics: Chamfer distance Hausdorff distance normal consistency precision recall F-score
Baselines: Poisson Voronoi POCO PCP RFEPS GeoUDF NKSR
Builds on: EC-Net DEF Poisson reconstruction RFEPS
Closest papers by meaning
- Preserving Topological and Geometric Embeddings for Point Cloud Recovery (AAAI 2026 · similarity 0.88)
- Deep Point Cloud Reconstruction (ICLR 2022 · similarity 0.88)
- LAM3D: Large Image-Point Clouds Alignment Model for 3D Reconstruction from Single Image (NeurIPS 2024 · similarity 0.88)
- Point2CAD: Reverse Engineering CAD Models from 3D Point Clouds (CVPR 2024 · similarity 0.87)
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · similarity 0.87)
- SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (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.