NeuroNURBS: Learning Efficient Surface Representations for 3D Solids

TL;DR

The evaluation in solid generation and segmentation tasks indicates that the NeuroNURBS performs comparably and, in some cases, superior to UV-grids, but with a significantly improved efficiency: for training the surface autoencoder, GPU consumption is reduced by 86.7%; memory requirement drops by 79.9% for storing 3D solids.

Abstract

Boundary Representation (B-Rep) is the de facto representation of 3D solids in Computer-Aided Design (CAD). B-Rep solids are defined with a set of NURBS (Non-Uniform Rational B-Splines) surfaces forming a closed volume. To represent a surface, current works often employ the UV-grid approximation, i.e., sample points uniformly on the surface. However, the UV-grid method is not efficient in surface representation and sometimes lacks precision and regularity. In this work, we propose NeuroNURBS, a representation learning method to directly encode the parameters of NURBS surfaces. Our evaluation in solid generation and segmentation tasks indicates that the NeuroNURBS performs comparably and, in some cases, superior to UV-grids, but with a significantly improved efficiency: for training the surface autoencoder, GPU consumption is reduced by 86.7%; memory requirement drops by 79.9% for storing 3D solids. Moreover, adapting BrepGen for solid generation with our NeuroNURBS improves the FID from 30.04 to 27.24, and resolves the undulating issue in generated surfaces.

Figures (6)

• Figure 1: Visual comparison of surfaces generated by BrepGen (a) and our method NURBS-Gen (b). Our NURBS-Gen can directly generate valid NURBS parametrization and hence ensure the smoothness and regularity of the surfaces, in contrast to the BrepGen surfaces, which appear undulating.
• Figure 2: Diagram for NeuroNURBS. Right: two parts of NeuroNURBS, preprocessing and autoencoding NURBS parameters. Left: a simplified diagram for UV-grid, where the approximation from UV-grid back to NURBS surface is not deterministic.
• Figure 3: Qualitative evaluation. It is observe that DeepCAD often generates invalid (i.e., multiple and collapsed) solids; while BrepGen and our NURBS-Gen are able to synthesize clean and upstanding B-Rep solids. In \ref{['fig:Nurbs-Gen_qualitative_zoom']}, we visualize generated samples from BrepGen and NURBS-Gen with zooming-in for a deep-diving qualitative evaluation.
• Figure 4: More examples generated with NURBS-Gen.
• Figure 5: NURBS-GAT diagram.