EXPLICIT ANNOTATED 3D-CNN DEEP LEARNING OF GEOMETRIC PRIMITIVES INSTANCES

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