Handling of “unknown unknowns” - classification of 3D geometries from CAD open set datasets using Convolutional Neural Networks

Abstract

This paper refers to the application of Convolutional Neural Networks (CNNs) for the classification of 3D geometries from Computer-Aided Design (CAD) datasets with a large proportion of unknown unknowns (classes unknown after training). The motivation of the work is the automatic recognition of standard parts in the large CAD-based image data set and thus, reducing the time required for the manual preparation of the data set. The classification is based on a threshold value of the Softmax output layer (first criterion), as well as on three different methods of a second criterion. The three methods for the second criterion are the comparison of metadata relating to the geometries, the comparison of feature vectors from previous dense layers of the CNN with a Spearman correlation, and the distance-based difference between multivariate Gaussian models of these feature vectors using Kullback-Leibler divergence. It is confirmed that all three methods are suitable to solve an open set problem in large 3D datasets (more than 1000 different geometries). Classification and training are image-based using different multi-view representations of the geometries.