Learn the Solid, Not the File: Canonical Inputs for Neural Networks on CAD Boundary Representations
Heinrich Jiang, Hager Yasser Mohamed, Alexander Hitt, Valeriia Lomakina, Henning Jiang, Jennifer Jang
cs.CV · cs.AI · cs.CG
Quick verdict
Proposes canonical region graph giving invariance guarantees for B-rep nets; strong evidence of baseline collapse, but no code URL is stated.
Summary
The paper shows that popular B-rep encoders (UV-Net, BRepNet, AAGNet) collapse when the same solid arrives in a different but equally valid boundary representation — after face re-partitioning, rigid motion, NURBS re-expression, a Rhino import/export round-trip, or independent re-modeling by CAD experts (25 parts, 100 FreeCAD programs from two labelers). It proposes the canonical region graph: faces lying on one underlying analytic surface are merged via exact surface keys (NURBS-disguised primitives recovered by deterministic least-squares fitting), features are expressed in a solid-derived canonical frame (boundary centroid, principal axes, sign fixed by third moments with continuous blending), and the solid is scaled to unit boundary area. Propositions 1-4 and Lemma 1 prove invariance to repartitioning and rigid motions, exact for closed-form features and O(ε) for mesh-quadrature features; Appendix F verifies this at benchmark scale. On MFInstSeg, MFCAD++ and CADSynth segmentation the method holds ~0.985-0.992 mIoU under every perturbation while baselines fall as low as 0.0055 (AAGNet, composed, MFInstSeg); clean accuracy matches the strongest baseline. In retrieval on 3,000 Fusion 360 parts it is the only model consistent across perturbations (96.1% axis-split) though below BRepNet under NURBS (80.5% vs 98.1%). Predictive churn is near zero. Limitations: partition invariance costs accuracy on Fusion 360 operation labels (0.6148 vs AAGNet's 0.7412 clean); the human study is small; no code URL is stated.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Existing B-rep encoders (UV-Net, BRepNet, AAGNet) collapse under valid same-solid B-rep variations, e.g. MFInstSeg AAGNet drops 0.9851→0.0604 under NURBS re-expression (Table 2).
- The canonical region graph is provably invariant to face repartitioning and rigid motions (Propositions 1–4, Lemma 1), verified at benchmark scale in Appendix F.
- The method matches the strongest baselines on clean machining benchmarks (~0.985–0.992 mIoU) and remains stable under all tested perturbations (Table 2).
- Architecture-level invariance outperforms per-perturbation data augmentation, which transfers poorly across perturbation types (Table 5).
- Partition-invariant predictors are bounded by geometry-only Bayes risk (Proposition 3), realized as the Fusion 360 clean deficit (0.6148 vs AAGNet 0.7412, Table 9).
Strengths
Real-world threat model, not only synthetic attacks
Beyond synthetic splits (Table 2), the paper runs a plain Rhino import/export over the entire MFInstSeg test split (Table 7: AAGNet 0.9851→0.2838, UV-Net 0.9725→0.0433) and a human study of independently authored FreeCAD programs (Tables 8, 10), grounding the problem in actual CAD workflows.
Theory matched to measurement
Propositions 1–4 and Lemma 1 are complemented by Appendix F representation-level verification on 3,000 MFInstSeg solids: 99.9–100% identical region graphs, exact-class worst residual 0.002σ, with an honest split between proof-invariant features and mesh-tolerance features that carry no guarantee.
Clean attribution of the gains
Transparent experimental bookkeeping
Weaknesses
No variance or significance reporting
No code or data release URL
Residual non-invariance at the embedding level
Small, single-system human study
Weak invariance contrast baseline
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets representation robustness in learning on CAD boundary representations (B-reps): the same solid admits many valid B-reps (different face partitionings, coordinate frames, analytic-vs-NURBS surface encoding, kernel translation), and encoders trained on the file's incidental choices fail when those choices change. Central claims: (1) UV-Net, BRepNet and AAGNet collapse under same-solid B-rep variation; (2) the proposed canonical region graph — nodes are maximal connected same-surface regions, features live in a solid-derived canonical frame, scale is normalized to unit boundary area — is provably invariant to repartitioning and rigid motions (Propositions 1–4, Lemma 1); (3) it matches the strongest baseline on clean machining benchmarks while staying stable under every tested perturbation.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is a canonical region graph?
An input representation for neural networks on CAD B-rep models in which faces lying on the same underlying surface (e.g., several arcs of one cylinder) are merged into single region nodes, and all geometric features are computed in a coordinate frame derived from the solid itself (boundary centroid plus principal axes) rather than the file's arbitrary frame. The result is invariant to how the solid happens to be partitioned into faces and to rigid motions.
Why do B-rep neural networks fail on the same part exported from different CAD software?
Because their features and topological kernels read file-level details — face partitioning, UV parameterizations, coordinate frames — that change on import/export or when a geometry kernel re-saves the file. The paper measures this: after a plain Rhino round-trip of the MFInstSeg test set, AAGNet segmentation mIoU falls from 0.9851 to 0.2838 and UV-Net from 0.9725 to 0.0433, while the proposed method stays at 0.9870.
Does the proposed method beat UV-Net, AAGNet, and BRepNet?
On clean machining benchmarks it matches them (0.9862 vs AAGNet's 0.9851 mIoU on MFInstSeg). Under B-rep perturbations it vastly outperforms, holding roughly 0.985–0.992 mIoU while baselines collapse to as low as 0.0055. One exception: on Fusion 360 operation-label segmentation with clean data it is worse (0.6148 vs AAGNet's 0.7412), because invariance deliberately discards modeling-history information.
Is code or data available for this paper?
The paper text contains no code or data release URL. Appendix B refers to released checkpoints without giving a location, and no repository link is stated. The construction is documented in unusual detail (tolerances, constants, protocol in Appendices E–F), but independent reproduction would require reimplementing the pipeline from that description.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Mathematical notation in the Appendix A proofs is garbled by PDF extraction (integral and expectation symbols); proofs were assessed at face value from the extracted wording.
- No code/data URL appears in the extracted text; Appendix B's 'released checkpoints' phrasing is ambiguous about what is actually available and where.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.