Abstract
Computer-aided design (CAD) is ubiquitous—virtually every modern object is designed using editable CAD tools. However, the availability of CAD datasets in their native, parametric boundary representation (B-Rep) format remains limited. This scarcity underscores the growing need for data-efficient learning methods in this domain.
In this paper, we introduce Masked Topology Modeling (MTM), a novel self-supervised pretraining task that leverages the face-adjacency graph—a structural feature unique to B-Rep representations. The encoder is tasked with reconstructing masked edges: specifically, a fraction of edges are hidden, and a small prediction head is trained to classify each masked edge's convexity and curve type based on the encoder's post-message-passing face features.
To further enhance learning, MTM is combined with:
- A MoCo-style momentum-queue contrastive learning framework, adapted with B-Rep-aware augmentations,
- A BFS-connected face-region masked reconstruction objective,
- Pretraining on the ABC dataset and a newly developed procedurally generated dataset.
This combination achieves strong performance across several benchmarks, demonstrating the effectiveness of MTM for self-supervised learning on parametric CAD.
Subjects: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
Cite as: arXiv:2607.20642 [cs.CV]
Submission history:
- [v1] Wed, 22 Jul 2026 18:13:08 UTC (546 KB)
Full-text links: Access the paper via arXiv.
via ArXiv CV