Brep2Shape: Boundary and Shape Representation Alignment via Self-Supervised Transformers
Researchers have developed Brep2Shape, a new self-supervised pre-training method designed to align boundary representations with shape representations in computer-aided design (CAD). This method addresses the representation gap between continuous and discrete approaches by learning to predict spatial points from Bézier control points. A Dual Transformer backbone with topology attention is employed to capture geometric properties and maintain topological consistency, leading to improved accuracy and faster convergence on downstream tasks. AI