Research map / Point clouds / Point cloud classification
SO(3)-Equivariant Learning on CAD Boundary Representations
Matteo Ballegeer, Dries Benoit
Read the paperCheck your idea against 137,217 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 an SO(3)-equivariant graph learning approach for CAD B-rep data that encodes surface geometry as equivariant features rather than rotation-sensitive scalars.
Problem
Existing B-rep CAD learning methods encode coordinates and normals as scalars, making them rotation-sensitive, while invariant alternatives based on per-face local frames lose a shared orientation reference needed to model inter-face directional relationships.
Approach
Uses SO(3)-equivariant learning on the B-rep geometric graph so that directional face-to-face relationships are captured under a shared global orientation reference, instead of scalar encodings or per-face local frames. Full method details are not available beyond the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SO(3)-equivariant CG message passing instead of FoV-Net's invariant local-frame messages | SO3brep · Fusion 360 segmentation | accuracy: 91.76 -> 95.11 | Improved |
| replace CG tensor product with non-equivariant message functions (ablation) | SO3brep · Fusion 360 SO(3) test | mIoU: 82.05 -> 61.89 (attention) | Hurt |
| equivariant instead of invariant message passing at full training data (ablation) | SO3brep · Fusion 360 / MFTRCAD | mIoU: +1.1 to +1.5 mIoU with 40% fewer parameters | Improved |
| equivariant instead of invariant message passing with 100 training parts (ablation) | SO3brep · MFTRCAD SO(3) test | mIoU: 29.87 -> 43.01 | Improved |
| non-equivariant messages on aligned parts at full data (ablation) | SO3brep · Fusion 360 original orientation | accuracy: 95.11 vs 95.31 (attention) | No clear effect |
| remove node features (ablation) | SO3brep · Fusion 360 segmentation | accuracy: -17.4 accuracy, -29.3 mIoU when removed | Improved |
| add edge features (ablation) | SO3brep · Fusion 360 segmentation | mIoU: -3.1 accuracy, -9.6 mIoU when removed | Improved |
| add attention aggregation instead of sum (ablation) | SO3brep · Fusion 360 segmentation | mIoU: -0.8 accuracy, -2.1 mIoU when removed | Improved |
| increase maximum irrep degree from 0 to 1 and 2 (ablation) | SO3brep · Fusion 360 segmentation | mIoU: -1.0 mIoU at lmax=1, -3.2 at lmax=0 | Improved |
| add equivariance to a pose-sensitive classification dataset | SO3brep · SolidLetters original orientation | accuracy: 97.36 (UV-Net) vs 97.03 | Hurt |
| SO3brep versus FoV-Net on BenDFM | SO3brep · BenDFM SO(3) test | accuracy: 81.13 -> 81.68 | No clear effect |
| CG tensor products add computational cost | SO3brep · Fusion 360 inference | inference time (ms): 8.3 -> 22.0 (vs FoV-Net) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- UV-Net — UV-Net: Learning From Boundary Representations (CVPR 2021) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- EquiCAD: A Geometric Equivariant Neural Network for 3D Shape Classification (ICML 2026 · similarity 0.92)
- FoV-Net: Rotation-Invariant CAD B-rep Learning via Field-of-View Ray Casting (CVPR 2026 · similarity 0.90)
- 3D Equivariant Graph Implicit Functions (ECCV 2022 · similarity 0.89)
- Learning Generalizable Shape Completion with SIM(3) Equivariance (NeurIPS 2025 · similarity 0.87)
- Grounding Continuous Representations in Geometry: Equivariant Neural Fields (ICLR 2025 · similarity 0.86)
- Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation (NeurIPS 2024 · similarity 0.86)
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.
