BRepCLIP: Contrastive Multimodal Pretraining on BRep Primitives for CAD Understanding
TLDR
BRepCLIP aligns BRep geometry with language and images via contrastive pretraining, improving CAD retrieval and classification.
Reasoning
The paper introduces a novel framework for learning BRep representations with multimodal alignment, achieving significant improvements in retrieval and zero-shot classification. However, it focuses on discriminative tasks and does not address generative modeling or reconstruction.
Read-first score
Read-first score 52.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 79.
Field roles
Rank sensitivity
Stability: volatile; rank range: 27.
Keyword Scores
Deep Analysis
Innovations
- First framework to align BRep geometry with language and image embeddings through contrastive pretraining
- Models CAD objects as sequences of face and edge tokens with separate discrete vocabularies for surface and curve geometry, augmented with spatial and semantic descriptors
- Transformer encoder aggregates tokens into a global BRep embedding aligned with CLIP's text and image encoders via a joint contrastive objective
- Generates more discriminative and semantically grounded embeddings than point-based alternatives
Methodology
BRepCLIP represents each CAD object as a sequence of face and edge tokens using separate discrete vocabularies for surface and curve primitives, enriched with spatial and semantic descriptors. A transformer encoder produces a global BRep embedding, which is aligned with CLIP's text and image embeddings through a joint contrastive loss.
Key Results
BRepCLIP improves Top-1 retrieval over OpenShape by 40.4%, 22.0%, and 23.9% on ABC, CADParser, and Automate, respectively, and boosts zero-shot classification Top-1 on FabWave by 15%. It also provides a CAD-aware similarity metric for evaluating text- and image-conditioned CAD generation.