Explicit Annotated 3D-CNN Deep Learning of Geometric Primitives Instances

DS 122: Proceedings of the Design Society: 24th International Conference on Engineering Design (ICED23)

Year: 2023
Editor: Kevin Otto, Boris Eisenbart, Claudia Eckert, Benoit Eynard, Dieter Krause, Josef Oehmen, Nad├Ęge Troussier
Author: Hilbig, Arthur; Holtzhausen, Stefan; Paetzold, Kristin
Series: ICED
Institution: Technische Universit├Ąt Dresden, Chair of Virtual Product Development
Section: Design Methods
Page(s): 1775-1784
DOI number: https://doi.org/10.1017/pds.2023.178


In reengineering technical components, the robust automation of reverse engineering (RE) could overcome the need for human supervision in the surface reconstruction process. Therefore, an enhanced computer-based geometric reasoning to derive tolerable surface deviations for reconstructing optimal surface models would promote a deeper geometric understanding of RE downstream processes. This approach integrates advanced surface information into a deep learning-based recognition framework by explicitly labeling geometric outliers and subsurface boundaries. For this purpose, a synthetic dataset is created that morphs nominal surface models to resemble the macroscopic surface pattern of physical components. For the detection of regular geometry primitives, a 3D-CNN is used to analyze the voxelized components based on signed distance field data. This explicit labeling approach enables surface fitting to derive suitable shape features that fulfill the underlying surface constraints.

Keywords: Reverse Engineering, Surface Reconstruction, Machine learning, Computer Aided Design (CAD), Artificial intelligence

Please sign in to your account

This site uses cookies and other tracking technologies to assist with navigation and your ability to provide feedback, analyse your use of our products and services, assist with our promotional and marketing efforts, and provide content from third parties. Privacy Policy.