Pith. sign in

REVIEW 4 major objections 4 minor 153 references

A new self-supervised objective, Masked Topology Modeling, hides edges of a CAD solid's face-adjacency graph and predicts their convexity and curve type, yielding label-efficient encoders that outperform prior methods on several B-rep bench

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 09:46 UTC pith:GY3ZNE4J

load-bearing objection Genuinely new MTM objective with solid few-shot gains; SOTA claim overstated, and the zero-loss theory needs an extra resolution assumption. the 4 major comments →

arxiv 2607.20642 v1 pith:GY3ZNE4J submitted 2026-07-22 cs.CV cs.LG

Masked Topology Modeling for Self-Supervised Learning on Parametric CAD

Heinrich Jiang , Jennifer Jang This is my paper
classification cs.CV cs.LG
keywords self-supervised learningCADB-repface-adjacency graphmasked topology modelingconvexity predictionfew-shot learninggraph neural networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper aims to establish a new self-supervised objective for boundary-representation (B-rep) CAD models: mask edges of the face-adjacency graph—the structure that records which faces meet along shared curves—and train a small head to predict each hidden edge's convexity and curve type, labels extracted for free by the geometry kernel. The authors argue this is the first masked graph-topology prediction for B-reps, and that it forces the encoder to represent the relative fold direction of adjacent faces, information that face-level methods (masked face reconstruction, face contrastive learning) provably ignore. They combine MTM with momentum-queue contrastive learning, a connected-region masked reconstruction objective, and pretraining on a public CAD corpus plus a procedurally generated dataset. Empirically, they report state-of-the-art label-efficient performance on several public B-rep benchmarks, with especially large gains in few-shot regimes (0.1%–1% labels). If correct, MTM offers a reusable, label-free pretraining signal that could help overcome the scarcity of editable CAD data.

Core claim

On its own terms, the paper claims that the convexity of a masked edge is always learnable to perfect accuracy from the endpoint faces' sampled geometry, whereas curve type is not, because different curves (e.g., a spline and a circle) can coincide at all finite sample points (Theorem 1). It further claims a separability result (Theorem 2): there exist two solids that are indistinguishable to any face-level representation—same face shapes, same adjacency graph—yet differ in a convexity label, and MTM must distinguish them while face-level objectives may merge them. The method combines MTM with momentum-queue contrastive learning, a connected-region masked reconstruction objective, and pretra

What carries the argument

The central object is the face-adjacency graph of a B-rep solid: nodes are faces, and an edge joins two faces that meet along a shared B-rep edge. Each graph edge carries two kernel-computed labels—convexity (convex/concave/smooth/knife) and curve type (line/circle/ellipse/B-spline/other)—which are free by construction. MTM deletes a random 70% of graph edges before message passing, so the encoder never sees the hidden edge, and trains two small MLP heads to predict those labels from the post-message-passing features of the two endpoint faces. The theoretical machinery is an identifiability proof (convexity ceiling zero, curve-type ceiling positive) and a separability proof (bump vs. dimple

Load-bearing premise

The proof that convexity is always learnable rests on the assumption that each face's finite sample grid resolves the outward normal field right next to every masked edge; on real CAD faces trimmed by curved boundaries, the standard 10x10 grid may have no sample that close to the seam, so the fold direction would not be fully determined.

What would settle it

Find a B-rep solid with a convex edge whose two faces meet at a small angle, and construct a UV grid that places no sample point within, say, 5% of the edge's maximum parameter range of that seam; if a pretrained MTM encoder's convexity head cannot classify such edges above chance, or if the convexity cross-entropy loss fails to approach zero as the grid is refined, then Assumption A3 is violated and the identifiability proof does not apply to practical sampling.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • B-rep encoders can be pretrained entirely on unlabeled or synthetic CAD solids and then transferred to labeled tasks with very few examples, greatly reducing annotation cost.
  • Predicting masked topological relations (rather than reconstructing masked geometry) is a viable discriminative self-supervised signal for graph-structured geometric data.
  • The theoretical ceiling on curve-type prediction suggests that coarser or higher-resolution sampling changes what is learnable, giving a guide for designing future SSL targets.
  • The synthetic-data generator provides a scalable source of topology-supervised B-rep solids, which the ablations show contributes substantial gains over real data alone.
  • If the separability result holds, MTM-like objectives are strictly more discriminative than any face-only representation for distinguishing local folds, which matters for tasks like machining-feature recognition.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The curve-type ceiling is a measurable prediction: as UV-grid resolution increases, the type head's loss floor should drop; this could be tested by pretraining with denser grids and comparing the type-loss plateau.
  • The method's principle—mask a relational attribute and predict it from neighbor context—may transfer to other structured 3D representations (meshes with sharp features, point clouds with adjacency) where a kernel or geometric measure can supply free labels.
  • Convexity prediction could be refined into regression on the dihedral angle, which the paper lists as future work; if that works, the representation would carry continuous fold information rather than four coarse classes.
  • The reported few-shot gains are mostly attributable to pretraining, so one could ablate whether the contrastive queue or the topological head matters more at extreme label scarcity; the paper's ablations isolate MTM but not the full decomposition at 0.1% labels.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper introduces Masked Topology Modeling (MTM), a self-supervised objective for B-rep face-adjacency graphs: a fraction of graph edges are deleted and a lightweight head predicts each deleted edge's convexity (4 classes) and curve type (5 classes) from the endpoint face embeddings produced by a UV-Net-style encoder after message passing. MTM is combined with MoCo contrastive learning, BFS-connected region masked reconstruction, and a pretraining corpus of ABC plus a procedurally generated dataset. The authors report large few-shot gains on F360, SolidLetters, MFInstSeg, and CADSynth relative to their baselines, and they give two theoretical results: convexity is identifiable to zero loss while curve type is not (Theorem 1), and any zero-MTM-loss representation must separate a bump/dimple pair that face-level representations may merge (Theorem 2).

Significance. If the claims are taken at face value, MTM is a useful, label-free pretraining signal for B-rep representation learning. The controlled MTM on/off ablation, the compute-matched synthetic-data ablation, and the consistent label-efficiency gains across four benchmarks are real strengths; the kernel-derived labels and detailed experimental protocols aid reproducibility. The main contributions are empirical, and those results seem credible. The theoretical claims, however, are more fragile, and one stated SOTA claim is contradicted by the paper's own table.

major comments (4)
  1. [§1, contribution 5; Table 2] The contributions list claims 'SOTA performance on ... SolidLetters ...', but Table 2 reports full-data 26-way accuracy 98.02±0.06 vs BRep-BERT's 98.54, and §5.1 states the full-data result is slightly worse. The SOTA claim is only true for the 10-way few-shot setting. Please qualify the claim or remove 'SOTA' for full-data SolidLetters.
  2. [§4, Assumption A3 (Appendix A/F)] A3 asserts the fixed UV-grid samples 'resolve the normal near σ_e' and calls this 'matching real encoders.' In the actual encoder, each face is sampled on a 10×10 grid over the parameter domain, and a B-rep edge is a trim curve that generally crosses the grid between samples; the masked edge's own curve samples are deleted, so the endpoint face grids are the only evidence. Nothing guarantees a sample lies in any neighborhood of σ_e, so A3 is not satisfied by construction. Theorem 1's convexity zero-loss result and Theorem 2's 'any zero-MTM-loss representation must separate' both rely on A3. If A3 fails, those conclusions do not apply to the implemented model. The empirical ablation stands, but the theoretical claim is overbroad. Please add a targeted experiment (e.g., convexity accuracy of the MTM head vs a high-resolution edge-aligned sampling control) or revise the assumptions/architec
  3. [§4, Theorem 1 proof (Appendix D)] The lower-bound argument for curve type writes (h_i,h_j)=g(S_e), using the claim that 'each h_v is a function of face v's finite sample set.' In the implemented encoder (Appendix A), h_v is produced after GATv2 message passing and global self-attention over all face tokens, so h_v depends on samples from many other faces, not only face v's 10×10 grid. Thus E[H(t(e)|h_i,h_j)] ≥ E[H(t(e)|S_e)] does not follow; a global encoder can exploit other faces' geometry to reduce residual type error below the local bound. The conclusion may still hold, but the proof as written establishes only a lower bound for local encoders. Please restrict the theorem or prove the bound for the actual architecture.
  4. [§5.2 vs Appendix E (Figure 4)] The text says the MTM on/off ablation trains both 9.4M and 27.9M models 'with MoCo and region masking enabled in both arms, varying only λ_topo (0.5 vs. 0.0).' Appendix E lists the 9.4M testbed with MoCo off and λ_topo=0.3 (and region ratio 0.4). This makes the description of the controlled comparison inaccurate: for the 9.4M rows, MoCo is not enabled, and the on-arm weight differs from 0.5. Please align the text and table, or report the exact setup used.
minor comments (4)
  1. [§3.5 vs Appendix E] The headline is described as a '30-epoch continuation from an ABC+synth-pretrained checkpoint' but Appendix E lists 'Epochs (from scratch) 30'. Clarify the pretraining schedule.
  2. [Table 4] The sentence 'our best seed (94.64) exceeding both' uses a best-seed number not shown in the table; report the mean-only comparison or provide the full seed distribution.
  3. [Figure 4] The y-axis labels appear garbled ('mIoUm' / 'mIoUw'); fix the typography.
  4. [Throughout] Use consistent capitalization for HierMAE/HierMae and BRep-BERT.

Circularity Check

0 steps flagged

No material circularity: MTM targets are kernel-derived and masked, the theory is conditional, and the only self-citation is background.

full rationale

MTM's self-supervised targets (convexity and curve type) are computed deterministically by the CAD kernel and are masked from the encoder before message passing; the loss in Eq. (1) is computed from endpoint face features only, so the objective is not fitted to downstream benchmarks and no fitted parameter is renamed as a prediction. The theoretical claims in Section 4 are conditional rather than circular: Lemma 1 derives convexity from the definition of dihedral angle and the assumed normal samples (A3); this is near-tautological but does not assume the conclusion. The possible mismatch between A3's 'matching real encoders' and the fixed 10x10 UV grid in Appendix A is an assumption/rigor gap, not a circular dependency. Theorem 2's bump/dimple construction follows directly from the definition of the MTM task and does not reuse an input as an output. The only self-citation (Bahri et al., 2022, Scarf) appears in introductory background and is not load-bearing. The empirical contribution is checked against external benchmarks with from-scratch controls and an MTM on/off ablation, so the central claims are self-contained.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The central claim rests on hand-set pretraining hyperparameters and augmentation schedules, the domain assumption that kernel-computed labels are reliable, and the stated assumptions of the two theorems. No new physical entities are introduced. The most fragile ledger item is the validity of the Theorem 2 witness construction, which is used to support the separability claim.

free parameters (8)
  • MTM mask rate p_mask = 0.7
    Fraction of graph edges masked before encoding; chosen by hand and used in both headline and ablation runs.
  • MTM loss weight lambda_topo = 0.5 (headline), 0.3 (9.4M testbed)
    Weight of the topology cross-entropy terms in the combined loss; not swept in the paper.
  • Region loss weight and region ratio = lambda_region=1.0, r=0.4
    Weight of BFS-region reconstruction and the fraction of faces masked in the connected region.
  • MoCo queue size, momentum, temperature = K=16,384, m=0.999, tau=0.07
    Contrastive dictionary hyperparameters taken from MoCo and hand-set.
  • Offline perturbation counts and operator probabilities = n_perts=5, min_perts=2, 14-operator table
    Distribution over B-rep operators (drill, boss, fillet, etc.) used to construct positive pairs; probabilities are chosen, not learned.
  • Online augmentation rates = p_face=0.3, p_edge=0.3, subgraph=0.5/0.7, jitter=0.005
    Face drop, edge drop, connected-subgraph sampling, and coordinate jitter rates; hand-set.
  • Convexity classification thresholds = n_a dot n_b > 0.999 smooth, < -0.999 knife
    Hand-set thresholds mapping the signed dihedral angle to four convexity buckets; affects the label distribution of the MTM target.
  • Backbone architectural dimensions = H=384, GATv2 layers=4, heads=6, Transformer layers=8, heads=6, FFN=1536
    Architecture size chosen to match BRep-BERT's scale (~22M); the 9.4M ablation uses different width/depth. These are hand-chosen, not derived.
axioms (5)
  • domain assumption B-rep geometry-kernel labels (convexity and curve type) are deterministic and consistent for every edge.
    Section 3.1 and Appendix F rely on OpenCASCADE to compute signed dihedral angle and curve type without human annotation.
  • domain assumption Assumptions A1-A4 of Theorem 1: faces are C^1 near the edge, edges meet exactly one curve, encoders are finite-resolution, and ambiguous curve types occur with positive probability.
    These are stated conditions for the identifiability theorem; A3 in particular is load-bearing for the claim that convexity is perfectly learnable.
  • ad hoc to paper The witness solids in Theorem 2 are valid closed B-rep solids.
    Appendix D constructs S' by subtracting 'the open right pyramid'; as written this leaves the base square Q on the top face and creates a non-manifold T-junction, so validity is not established.
  • standard math Standard information-theoretic facts: cross-entropy is minimized by the conditional entropy, and the data-processing inequality holds.
    Used in the proofs of Theorems 1 and 2 in Appendix D.
  • domain assumption Pretraining on ABC plus the procedurally generated corpus transfers to the downstream F360, SolidLetters, MFInstSeg, and CADSynth benchmarks.
    The empirical method is out-of-distribution for F360 and MFInstSeg; the paper assumes the self-supervised features are general enough to transfer.

pith-pipeline@v1.3.0-alltime-deepseek · 18224 in / 22373 out tokens · 209489 ms · 2026-08-01T09:46:48.031210+00:00 · methodology

0 comments
read the original abstract

Computer aided design (CAD) is ubiquitous: virtually any modern object was designed using editable CAD tools. However, with the shortage of available CAD datasets in its native editable and parametric format, boundary representation (B-Rep), it is ever more important to develop data-efficient methods for this domain. We present a new self-supervised pretraining task, Masked Topology Modeling (MTM), that leverages the face-adjacency graph, an induced structure unique to B-reps that the encoder can be asked to reconstruct. MTM masks a fraction of edges and trains a small head to predict each masked edge's convexity and curve type from the encoder's post-message-passing face features. We combine MTM with a MoCo-style momentum-queue contrastive learning over B-rep-aware augmentations, a BFS-connected face-region masked-reconstruction objective, and pretraining on the ABC dataset and our new procedurally generated dataset to show strong performance on a number of benchmarks.

Figures

Figures reproduced from arXiv: 2607.20642 by Heinrich Jiang, Jennifer Jang.

Figure 1
Figure 1. Figure 1: A B-rep solid as an labeled face￾adjacency graph. Each face of the solid (left) be￾comes a node (right), and two nodes are joined by a graph edge iff their faces meet along a shared B￾rep edge. Every graph edge carries the two labels MTM predicts: a convexity class (convex, con￾cave, smooth, knife) and a curve type (line, cir￾cle, B-spline, ellipse, other), both computed de￾terministically by the geometry … view at source ↗
Figure 2
Figure 2. Figure 2: Two-level positive-pair construction. Offline, each source solid is perturbed by B-rep op￾erators into several cached variants that differ in geometry. Online, a training pair draws two of these variants and augments them asymmetrically at the graph level. One receives heavy augmentation to form view A (the contrastive query), the other light augmentation to form view B (the contrastive key, and the graph … view at source ↗
Figure 3
Figure 3. Figure 3: Masked Topology Modeling (MTM). Left: each edge of the face-adjacency graph carries kernel-computed labels — a convexity class and a curve type. Middle: we hide one edge, deleting it and its labels from the graph before encoding, so no message crosses that seam. Right: from the two endpoint faces alone (each contributing only its own boundary evidence, never the removed edge) an MTM head recovers the hidde… view at source ↗
Figure 4
Figure 4. Figure 4: MTM On vs. Off: We plot results on frozen F360 linear probe metrics for both 9.4M and 27.9M models with MTM on vs off. We see that MTM on improves metrics across the board [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

153 extracted references · 21 linked inside Pith

  1. [1]

    Geometric modeling of solid objects by using a face adjacency graph representation

    Silvia Ansaldi, Leila De Floriani, and Bianca Falcidieno. Geometric modeling of solid objects by using a face adjacency graph representation. ACM SIGGRAPH Computer Graphics, 19 0 (3): 0 131--139, 1985

    1985

  2. [2]

    Scarf: Self-supervised contrastive learning using random feature corruption

    Dara Bahri, Heinrich Jiang, Yi Tay, and Donald Metzler. Scarf: Self-supervised contrastive learning using random feature corruption. ICLR, 2022

    2022

  3. [3]

    Geometric continuity of parametric curves: three equivalent characterizations

    Brian A Barsky and Tony D DeRose. Geometric continuity of parametric curves: three equivalent characterizations. IEEE Computer Graphics and Applications, 9 0 (6): 0 60--69, 1989

    1989

  4. [4]

    Automatic reconstruction of 3d cad models from digital scans

    Fausto Bernardini, Chandrajit L Bajaj, Jindong Chen, and Daniel R Schikore. Automatic reconstruction of 3d cad models from digital scans. International Journal of Computational Geometry & Applications, 9 0 (04n05): 0 327--369, 1999

    1999

  5. [6]

    Big self-supervised models are strong semi-supervised learners

    Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33: 0 22243--22255, 2020

    2020

  6. [7]

    Spelsnet: Surface primitive elements segmentation by b-rep graph structure supervision

    Kseniya Cherenkova, Elona Dupont, Anis Kacem, Gleb Gusev, and Djamila Aouada. Spelsnet: Surface primitive elements segmentation by b-rep graph structure supervision. Advances in Neural Information Processing Systems, 37: 0 1251--1269, 2024

    2024

  7. [8]

    Hierarchical cadnet: Learning from b-reps for machining feature recognition

    Andrew R Colligan, Trevor T Robinson, Declan C Nolan, Yang Hua, and Weijuan Cao. Hierarchical cadnet: Learning from b-reps for machining feature recognition. Computer-Aided Design, 147: 0 103226, 2022

    2022

  8. [9]

    Cadops-net: Jointly learning cad operation types and steps from boundary-representations

    Elona Dupont, Kseniya Cherenkova, Anis Kacem, Sk Aziz Ali, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Cadops-net: Jointly learning cad operation types and steps from boundary-representations. In 2022 International Conference on 3D Vision (3DV), pp.\ 114--123. IEEE, 2022

    2022

  9. [10]

    Computational geometry for design and manufacture

    Ivor D Faux and Michael J Pratt. Computational geometry for design and manufacture. Halsted Press, 1979

    1979

  10. [11]

    Computer-aided diagnosis: The emerging of three cad systems induced by japanese health care needs

    Hiroshi Fujita, Yoshikazu Uchiyama, Toshiaki Nakagawa, Daisuke Fukuoka, Yuji Hatanaka, Takeshi Hara, Gobert N Lee, Yoshinori Hayashi, Yuji Ikedo, Xin Gao, et al. Computer-aided diagnosis: The emerging of three cad systems induced by japanese health care needs. Computer methods and programs in biomedicine, 92 0 (3): 0 238--248, 2008

    2008

  11. [12]

    CAD/CAM: computer-aided design and manufacturing

    Mikell Groover and EWJR Zimmers. CAD/CAM: computer-aided design and manufacturing. Pearson Education, 1983

    1983

  12. [13]

    Complexgen: Cad reconstruction by b-rep chain complex generation

    Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Transactions on Graphics (TOG), 41 0 (4): 0 1--18, 2022

    2022

  13. [14]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020

    2020

  14. [15]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll \'a r, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16000--16009, 2022

    2022

  15. [16]

    Trends of 4d cad applications for construction planning

    David Heesom and Lamine Mahdjoubi. Trends of 4d cad applications for construction planning. Construction management and economics, 22 0 (2): 0 171--182, 2004

    2004

  16. [17]

    Uv-net: Learning from boundary representations

    Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G Lambourne, Karl DD Willis, Thomas Davies, Hooman Shayani, and Nigel Morris. Uv-net: Learning from boundary representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 11703--11712, 2021

    2021

  17. [19]

    Graph-based heuristics for recognition of machined features from a 3d solid model

    Sanjay Joshi and Tien-Chien Chang. Graph-based heuristics for recognition of machined features from a 3d solid model. Computer-aided design, 20 0 (2): 0 58--66, 1988

    1988

  18. [21]

    Intellectual property challenges in the age of 3d printing: Navigating the digital copycat dilemma

    Antreas Kantaros. Intellectual property challenges in the age of 3d printing: Navigating the digital copycat dilemma. Applied Sciences, 14 0 (23): 0 11448, 2024

    2024

  19. [22]

    Unveiling the shadows: Anticipating future cyber risks and their impacts on businesses

    C etin Karahan, Hakan Velio g lu, and Yenal Arslan. Unveiling the shadows: Anticipating future cyber risks and their impacts on businesses. In Futurisks: Risk management in the digital age, pp.\ 17--55. Springer, 2025

    2025

  20. [23]

    ABC : A big CAD model dataset for geometric deep learning

    Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. ABC : A big CAD model dataset for geometric deep learning. CVPR, 2019

    2019

  21. [24]

    Brepnet: A topological message passing system for solid models

    Joseph G Lambourne, Karl DD Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. Brepnet: A topological message passing system for solid models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 12773--12782, 2021

    2021

  22. [25]

    Mesh and geometry manipulations for optimization and inverse design

    Benoit Lecallard, Trevor T Robinson, and Simao P Marques. Mesh and geometry manipulations for optimization and inverse design. In AIAA Scitech 2021 Forum, pp.\ 1901, 2021

    2021

  23. [26]

    Enhancing cad data integrity and security in supply chain networks using blockchain

    Chengnan Li, Tong Yu, Wangyan Li, Ying Liu, and Hongjun Yang. Enhancing cad data integrity and security in supply chain networks using blockchain. International Journal of Information Systems and Supply Chain Management (IJISSCM), 18 0 (1): 0 1--22, 2025

    2025

  24. [28]

    Fast industrial product design method and its application based on 3d cad system

    Fei Liu. Fast industrial product design method and its application based on 3d cad system. Computer-Aided Design & Applications, 18, 2021

    2021

  25. [29]

    Brep-bert: Pre-training boundary representation bert with sub-graph node contrastive learning

    Yunzhong Lou, Xueyang Li, Haotian Chen, and Xiangdong Zhou. Brep-bert: Pre-training boundary representation bert with sub-graph node contrastive learning. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pp.\ 1657--1666, 2023

    2023

  26. [31]

    Masked autoencoders for 3d point cloud self-supervised learning

    Yatian Pang, Eng Hock Francis Tay, Li Yuan, and Zhenghua Chen. Masked autoencoders for 3d point cloud self-supervised learning. World Scientific Annual Review of Artificial Intelligence, 1: 0 2440001, 2023

    2023

  27. [33]

    Cad-recode: Reverse engineering cad code from point clouds

    Danila Rukhovich, Elona Dupont, Dimitrios Mallis, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad-recode: Reverse engineering cad code from point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 9801--9811, 2025

    2025

  28. [34]

    Intellectual property protection

    Josip Stjepandi \'c , Harald Liese, and Amy JC Trappey. Intellectual property protection. In Concurrent engineering in the 21st century: Foundations, developments and challenges, pp.\ 521--551. Springer, 2015

    2015

  29. [35]

    CAD principles for architectural design

    Peter Szalapaj. CAD principles for architectural design. Routledge, 2013

    2013

  30. [36]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  31. [37]

    Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G

    Karl D.D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic CAD construction. In SIGGRAPH, 2021

    2021

  32. [38]

    Joinable: Learning bottom-up assembly of parametric cad joints

    Karl DD Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G Lambourne, Armando Solar-Lezama, et al. Joinable: Learning bottom-up assembly of parametric cad joints. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 15849--15860, 2022

    2022

  33. [39]

    Aagnet: A graph neural network towards multi-task machining feature recognition

    Hongjin Wu, Ruoshan Lei, Yibing Peng, and Liang Gao. Aagnet: A graph neural network towards multi-task machining feature recognition. Robotics and Computer-Integrated Manufacturing, 86: 0 102661, 2024

    2024

  34. [40]

    Autobrep: Autoregressive b-rep generation with unified topology and geometry

    Xiang Xu, Pradeep Jayaraman, Joseph Lambourne, Yilin Liu, Durvesh Malpure, and Pete Meltzer. Autobrep: Autoregressive b-rep generation with unified topology and geometry. In Proceedings of the SIGGRAPH Asia 2025 Conference Papers, pp.\ 1--12, 2025

    2025

  35. [42]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling

    Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 19313--19322, 2022

    2022

  36. [43]

    Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training

    Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35: 0 27061--27074, 2022

    2022

  37. [44]

    Brepmfr: Enhancing machining feature recognition in b-rep models through deep learning and domain adaptation

    Shuming Zhang, Zhidong Guan, Hao Jiang, Xiaodong Wang, and Pingan Tan. Brepmfr: Enhancing machining feature recognition in b-rep models through deep learning and domain adaptation. Computer Aided Geometric Design, 111: 0 102318, 2024

    2024

  38. [45]

    CVPR , year=

    Momentum Contrast for Unsupervised Visual Representation Learning , author=. CVPR , year=

  39. [46]

    ICCV , year=

    Emerging Properties in Self-Supervised Vision Transformers , author=. ICCV , year=

  40. [47]

    Zhou, Jinghao and Wei, Chen and Wang, Huiyu and Shen, Wei and Xie, Cihang and Yuille, Alan and Kong, Tao , booktitle=

  41. [48]

    Hou, Zhenyu and Liu, Xiao and Cen, Yukuo and Dong, Yuxiao and Yang, Hongxia and Wang, Chunjie and Tang, Jie , booktitle=

  42. [49]

    arXiv preprint arXiv:1611.07308 , year=

    Variational Graph Auto-Encoders , author=. arXiv preprint arXiv:1611.07308 , year=

    Pith/arXiv arXiv

  43. [50]

    Wu, Rundi and Xiao, Chang and Zheng, Changxi , booktitle=

  44. [51]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Cad-recode: Reverse engineering cad code from point clouds , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  45. [52]

    Koch, Sebastian and Matveev, Albert and Jiang, Zhongshi and Williams, Francis and Artemov, Alexey and Burnaev, Evgeny and Alexa, Marc and Zorin, Denis and Panozzo, Daniele , journal=

  46. [53]

    and Pu, Yewen and Luo, Jieliang and Chu, Hang and Du, Tao and Lambourne, Joseph G

    Willis, Karl D.D. and Pu, Yewen and Luo, Jieliang and Chu, Hang and Du, Tao and Lambourne, Joseph G. and Solar-Lezama, Armando and Matusik, Wojciech , booktitle=. Fusion 360 Gallery: A Dataset and Environment for Programmatic

  47. [54]

    Computer methods and programs in biomedicine , volume=

    Computer-aided diagnosis: The emerging of three CAD systems induced by Japanese health care needs , author=. Computer methods and programs in biomedicine , volume=. 2008 , publisher=

    2008

  48. [55]

    arXiv preprint arXiv:2505.19713 , year=

    CAD-Coder: Text-to-CAD Generation with Chain-of-Thought and Geometric Reward , author=. arXiv preprint arXiv:2505.19713 , year=

    Pith/arXiv arXiv

  49. [56]

    International Design Engineering Technical Conferences and Computers and Information in Engineering Conference , volume=

    Cad-coder: An open-source vision-language model for computer-aided design code generation , author=. International Design Engineering Technical Conferences and Computers and Information in Engineering Conference , volume=. 2025 , organization=

    2025

  50. [57]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Neural 3d mesh renderer , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  51. [58]

    Advances in Neural Information Processing Systems , volume=

    Unique3d: High-quality and efficient 3d mesh generation from a single image , author=. Advances in Neural Information Processing Systems , volume=

  52. [59]

    Advances in Neural Information Processing Systems , volume=

    One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization , author=. Advances in Neural Information Processing Systems , volume=

  53. [60]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Pixel2mesh: Generating 3d mesh models from single rgb images , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  54. [61]

    arXiv preprint arXiv:2404.07191 , year=

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models , author=. arXiv preprint arXiv:2404.07191 , year=

    Pith/arXiv arXiv

  55. [62]

    1983 , publisher=

    CAD/CAM: computer-aided design and manufacturing , author=. 1983 , publisher=

    1983

  56. [63]

    Construction management and economics , volume=

    Trends of 4D CAD applications for construction planning , author=. Construction management and economics , volume=. 2004 , publisher=

    2004

  57. [64]

    2013 , publisher=

    CAD principles for architectural design , author=. 2013 , publisher=

    2013

  58. [65]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Clip2: Contrastive language-image-point pretraining from real-world point cloud data , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  59. [66]

    arXiv preprint arXiv:1810.05795 , year=

    Point cloud gan , author=. arXiv preprint arXiv:1810.05795 , year=

    Pith/arXiv arXiv

  60. [67]

    European Conference on Computer Vision , pages=

    Image2point: 3d point-cloud understanding with 2d image pretrained models , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  61. [68]

    Proceedings of the European Conference on Computer Vision (ECCV) Workshops , pages=

    Image-to-voxel model translation with conditional adversarial networks , author=. Proceedings of the European Conference on Computer Vision (ECCV) Workshops , pages=

  62. [69]

    European Conference on Computer Vision , pages=

    Image-to-voxel model translation for 3d scene reconstruction and segmentation , author=. European Conference on Computer Vision , pages=. 2020 , organization=

    2020

  63. [70]

    Advances in neural information processing systems , volume=

    Voxel-based 3D detection and reconstruction of multiple objects from a single image , author=. Advances in neural information processing systems , volume=

  64. [71]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    CADDreamer: CAD Object Generation from Single-view Images , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  65. [72]

    ACM Transactions on Graphics (TOG) , volume=

    Brepgen: A b-rep generative diffusion model with structured latent geometry , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  66. [73]

    LLaVA-NeXT: Improved reasoning, OCR, and world knowledge , url=

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Li, Bo and Zhang, Yuanhan and Shen, Sheng and Lee, Yong Jae , month=. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge , url=

  67. [74]

    arXiv preprint arXiv:2409.16294 , year=

    Gencad: Image-conditioned computer-aided design generation with transformer-based contrastive representation and diffusion priors , author=. arXiv preprint arXiv:2409.16294 , year=

    Pith/arXiv arXiv

  68. [75]

    arXiv preprint arXiv:2411.15279 , year=

    Don't Mesh with Me: Generating Constructive Solid Geometry Instead of Meshes by Fine-Tuning a Code-Generation LLM , author=. arXiv preprint arXiv:2411.15279 , year=

    Pith/arXiv arXiv

  69. [76]

    arXiv preprint arXiv:2307.14377 , year=

    How can large language models help humans in design and manufacturing? , author=. arXiv preprint arXiv:2307.14377 , year=

    Pith/arXiv arXiv

  70. [77]

    arXiv preprint arXiv:2308.00227 , year=

    Experiments on Generative AI-powered parametric modeling and BIM for architectural design , author=. arXiv preprint arXiv:2308.00227 , year=

    Pith/arXiv arXiv

  71. [78]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    Optimizing evolutionary csg tree extraction , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

  72. [79]

    Advances in neural information processing systems , volume=

    UCSG-NET-unsupervised discovering of constructive solid geometry tree , author=. Advances in neural information processing systems , volume=

  73. [80]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Csg-stump: A learning friendly csg-like representation for interpretable shape parsing , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  74. [81]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Capri-net: Learning compact cad shapes with adaptive primitive assembly , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  75. [82]

    Ieee Transactions on Pattern Analysis and Machine Intelligence , volume=

    Neural Shape Parsers for Constructive Solid Geometry , author=. Ieee Transactions on Pattern Analysis and Machine Intelligence , volume=

  76. [83]

    Journal of Nuclear Engineering , volume=

    A novel algorithm for CAD to CSG conversion in McCAD , author=. Journal of Nuclear Engineering , volume=. 2023 , publisher=

    2023

  77. [84]

    Nuclear Engineering and Technology , volume=

    GEOUNED: A new conversion tool from CAD to Monte Carlo geometry , author=. Nuclear Engineering and Technology , volume=. 2024 , publisher=

    2024

  78. [85]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    3d point cloud generative adversarial network based on tree structured graph convolutions , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  79. [86]

    International conference on learning representations , year=

    Learning localized generative models for 3d point clouds via graph convolution , author=. International conference on learning representations , year=

  80. [87]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Pointflow: 3d point cloud generation with continuous normalizing flows , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

Showing first 80 references.