Brep2Shape: Boundary and Shape Representation Alignment via Self-Supervised Transformers

Brep2Shape is a self-supervised pre-training method designed to align boundary representations (B-rep) with intuitive shape representations, addressing the representation gap in CAD models. It utilizes a Dual Transformer architecture with parallel streams for encoding surface and curve tokens, along with topology attention to maintain topological consistency. Experimental results indicate that Brep2Shape achieves state-of-the-art accuracy and faster convergence in downstream tasks, making it a significant advancement for practitioners working with geometric data in AI applications.