Research map / Point clouds / Point cloud classification
STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds
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 STAR-Edge, which projects each point's neighborhood onto a unit sphere to form a local spherical curve, encodes it with spherical harmonics into a rotation-invariant descriptor for MLP edge classification, and refines edge points using sphere-based normal estimation.
Problem
Extracting geometric edges from unstructured point clouds of thin-walled structures, where local neighborhoods are unreliable (mixing upper/lower surfaces, sparse or noisy side-end faces), degrading existing edge extraction methods.
Approach
Build a structure-aware local neighborhood by projecting KNN points onto a unit sphere centered at the query point; co-planar points align along a great circle arc, so the spherical curve emphasizes same-surface points. Encode it with spherical harmonic energies as a rotation-invariant descriptor for an MLP classifier, and reuse the representation to estimate normals (RANSAC great-circle) and proj
Key results
- On the thin-walled dataset, STAR-Edge achieves ECD 0.0587 and F1 0.7013 versus 0.4839 ECD for MFLE and 1.4392 for RFEPS; EC-Net and PIE-Net are worse (PIE-Net ECD 52.12).
- On the ABC dataset, STAR-Edge reports ECD 0.0024, the lowest among compared methods (NerVE 0.0026, RFEPS 0.0026, EC-NET 0.0037, PIE-NET 0.0090).
- Robustness: ECD stays around 0.059 for noise levels 0.001 and 0.005 and resolution 0.5; accuracy remains 0.993 across tested settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use STAR-Edge instead of MFLE | Thin-walled structure edge extraction · Thin-walled structure dataset | ECD: 0.4839 -> 0.0587 | Helped |
| use STAR-Edge instead of MFLE | Thin-walled structure edge extraction · Thin-walled structure dataset | F1: 0.4621 -> 0.7013 | Helped |
| use STAR-Edge instead of the strongest ECD baselines | Edge extraction · ABC dataset | ECD: 0.0026 -> 0.0024 | Helped |
| use SH bandwidth 10 instead of 5 (ablation) | STAR-Edge · Thin-walled structure dataset | ECD: 0.122 -> 0.059 | Helped |
| use SH bandwidth 10 instead of 5 (ablation) | STAR-Edge · Thin-walled structure dataset | F1: 0.685 -> 0.701 | Helped |
| use SH bandwidth 15 instead of 10 (ablation) | STAR-Edge · Thin-walled structure dataset | ECD: 0.059 -> 0.048 | Helped |
| use SH bandwidth 15 instead of 10 (ablation) | STAR-Edge · Thin-walled structure dataset | F1: 0.701 -> 0.785 | Helped |
Surprising findings
- PIE-Net achieves a very high ECD (about 52) on the thin-walled dataset despite reasonable accuracy, indicating that point-classification metrics alone can mislead for edge extraction.
- Precision stays low (about 0.54) for STAR-Edge on the thin-walled dataset, so accuracy near 0.99 is largely driven by the class imbalance between edge and non-edge points.
Limitations
- Running time (214 s on ~320k points) is much higher than several baselines.
- Performance degrades at high noise (0.03) and low resolution (0.8), where very few side-face points are sampled.
- Precision on the thin-walled dataset remains modest (about 0.54 at the default setting).
Experimental setup
Datasets: Thin-walled structure dataset [6] (APEE-Net) ABC dataset
Benchmarks: Thin-walled structure dataset [6] ABC dataset
Metrics: Recall Precision F1 Accuracy Edge Chamfer Distance (ECD)
Baselines: EC-Net PIE-Net NerVE MFLE RFEPS
Builds on: PIE-Net RFEPS EC-Net NerVE
Closest papers by meaning
- EdgeMovingNet: Edge-preserving Point Cloud Reconstruction via Joint Geometry Features (CVPR 2025 · similarity 0.86)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.85)
- CurveCloudNet: Processing Point Clouds with 1D Structure (CVPR 2024 · similarity 0.85)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.84)
- Structure from Collision (CVPR 2025 · similarity 0.84)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.84)
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.