Reproduce “DualBrep: A Dual-Field Continuous Representation for B-rep Modelling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yilin Liu, Pradeep Jayaraman, Chinthala Reddy, Xiang Xu, Hooman Shayani
DualBrep introduces a system that transforms Boundary Representation (B-rep) models into a unified dual-field continuous representation for deep learning. This enables robust reverse engineering and generative modeling of CAD shapes, achieving 76.34% topological validity in reconstruction and supporting conditional generation from point clouds and images.
Key insights
- Encoding B-rep geometry and topology into a unified continuous Euclidean domain resolves the "compatibility paradox" between discrete B-reps and gradient-based deep learning, enabling end-to-end optimization.
- The dual-field approach, combining SDF for geometry and a GVD-derived UDF for topology, provides a resolution-independent and differentiable representation that aids in generating globally consistent and watertight B-rep models.
- Training a conditional Flow Matching model over the VAE's latent space allows for scalable and controllable generative modeling from diverse inputs like point clouds or single-view images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.