Research map / Point clouds / Point cloud classification
SpelsNet: Surface Primitive Elements Segmentation by B-Rep Graph Structure Supervision
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 SpelsNet, a sparse-CNN point-cloud network that segments CAD point clouds into B-Rep edges/faces and types, with an extra graph head supervised by a point-to-BRep adjacency version of the Linear Algebraic Representation (LAR).
Problem
Segmenting 3D point clouds of CAD shapes into B-Rep surface primitive elements (edges, faces, their types and memberships) while keeping topologically consistent structure, without slow post-processing.
Approach
Transfers the LAR characteristic matrices of a B-Rep chain complex to the point cloud domain (point-to-BRep adjacency, LAR_pcd) and uses them as direct supervision for a learned graph structure (GSL layer) plus a GCN that predicts element types, jointly with spatial type and membership segmentation.
Key results
- On ABCParts-VEF with rotation augmentation, SpelsNet reaches 65.60 face sIoU and 81.93 face tIoU, versus 55.75/55.39 for PrimitiveNet.
- On CC3D-VEF with augmentation, SpelsNet achieves 21.23 sIoU and 45.15 tIoU, the best among compared methods.
- Topology error t_fe drops from 0.201 (ComplexGen) to 0.06 on ABCParts-VEF, with face type accuracy 90.01% vs 74.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SpelsNetvef joint topological supervision to spatial segmentation with coordinates and normals (ablation) | SpelsNet · ABCParts-VEF, w/aug, face | sIoU: 59.69 -> 65.60 | Helped |
| add point normals while omitting SpelsNetvef (ablation) | SpelsNet spatial model · ABCParts-VEF, w/aug, edge | sIoU: 45.47 -> 46.69 | Helped |
| add point normals while omitting SpelsNetvef (ablation) | SpelsNet spatial model · ABCParts-VEF, w/aug, face | sIoU: 58.53 -> 59.69 | Helped |
| use voxel resolution ρ=0.02 instead of ρ=0.01 (ablation) | SpelsNet · ABCParts, edge | sIoU: 50.01 -> 35.60 | Hurt |
| use voxel resolution ρ=0.02 instead of ρ=0.01 (ablation) | SpelsNet · ABCParts, face | tIoU: 81.93 -> 59.04 | Hurt |
| use voxel resolution ρ=0.005 instead of ρ=0.01 (ablation) | SpelsNet · ABCParts, edge | sIoU: 50.01 -> 43.76 | Hurt |
| use voxel resolution ρ=0.005 instead of ρ=0.01 (ablation) | SpelsNet · ABCParts, face | tIoU: 81.93 -> 61.96 | Hurt |
| use dynamic voxel density ψ=4-6 (ablation) | SpelsNet · ABCParts, edge | sIoU: 47.14 -> 51.54 | Helped |
| use dynamic voxel density ψ=4-6 (ablation) | SpelsNet · ABCParts, face | tIoU: 72.07 -> 83.74 | Helped |
| use SpelsNet instead of ComplexGen | B-Rep topology prediction · ABCParts-VEF | tfe: 0.201 -> 0.06 | Helped |
| use SpelsNet instead of PrimitiveNet | face segmentation · CC3D-VEF, face w/aug | sIoU: 15.22 -> 21.23 | Helped |
| use SpelsNet instead of ComplexGen | face type prediction · CC3D-VEF, face w/aug | tIoU: 19.66 -> 45.15 | Helped |
Surprising findings
- Performance of ParSeNet and HPNet drops significantly under random rotation augmentation, showing dataset alignment bias.
- The spatial head outperforms the topological GCN head, which the authors attribute to insufficient GCN capacity.
Limitations
- Performance degrades under rotation, and the topological GCN head underperforms the spatial head.
- Fixed characteristic matrix size (Ne=Nf=128) must be tuned per dataset.
- Sensitive to input voxel resolution; robustness requires dynamic resolution selection.
Experimental setup
Datasets: ABCParts CC3D
Benchmarks: ABCParts-VEF CC3D-VEF
Metrics: sIoU tIoU type accuracy t_fe topology error
Baselines: ParSeNet HPNet PrimitiveNet ComplexGen
Builds on: Linear Algebraic Representation (LAR) MinkowskiEngine SparseCNN ResUNet ComplexGen ParSeNet HDBSCAN
Closest papers by meaning
- BRepNet: A Topological Message Passing System for Solid Models (CVPR 2021 · similarity 0.87)
- SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation (NeurIPS 2025 · similarity 0.86)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.85)
- DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry (CVPR 2025 · similarity 0.84)
- Point2CAD: Reverse Engineering CAD Models from 3D Point Clouds (CVPR 2024 · similarity 0.84)
- BrepGiff: Lightweight Generation of Complex B-rep with 3D GAT Diffusion (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.