Reproduce “BRepNet: A topological message passing system for solid models” 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
Joseph G. Lambourne, Karl D.D. Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, Hooman Shayani
Researchers from Autodesk Research and UCL introduce BRepNet, a neural network architecture designed to directly process Boundary Representation (B-rep) 3D models, preserving their inherent geometric precision and topological information. The work also releases the Fusion 360 Gallery dataset of over 35,000 annotated B-rep models, with BRepNet achieving a 77.10% IoU for face segmentation, outperforming rule-based, graph neural network, and geometry-based deep learning approaches.
Key insights
- Leveraging the consistent local topology of B-rep coedges through specific, ordered convolutional kernels is crucial for accurate feature recognition and information propagation.
- Coordinate-free geometric features, particularly edge convexity, curve type, and surface type, are critical information sources for B-rep segmentation.
- Directly processing B-rep data with topology-aware convolutions offers superior performance and stability compared to general graph networks or deep learning on derived geometric approximations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.