Multi-view expressive graph neural networks for 3D CAD model classification

The creation of effective content-based retrieval systems for the 3D models of engineering components created by Mechanical Computer Aided Design (MCAD) systems has been a subject of academic investigation since the 1990 s. Recently some of the most promising results have been reported by researchers using Deep Neural Nets (DNN) to classify industrial parts represented by collections of images generated from different viewpoints. Known as Multi-View Convolutional Neural Network (MVCNN) these systems have extended the architectures developed for 2D images analysis to handle 3D data by using a series of pictures rendered from different viewpoints. In 2016 Monti et al. used Graph Neural Networks (GNN) to classify the superpixel images for the first time and reported results which suggested that the approach could produce good classification accuracy (an observation confirmed by the application of GNNs to 2D image datasets of common objects such as MNIST and Cifar). To investigate the potential for GNNs to reason about spatial adjacency relationships this paper reports, for the first time, the application of GNNs to classifications of 3D MCAD models of mechanical components. Viewing GNN as a convolution operator, parallel GNN can be applied to multiple views of a 3D shape in the same way as CNN is structured in a MVCNN. After outlining the architecture implemented to support MVGNNs the impact of various hyperparameters on the expressive power are explored. When optimised these hyper -parameters (located in both intra-layer and inter-layer structures) combined with the multi-view architecture produced a classification accuracy superior to previous methods. The results also suggest that GNNs have the potential to produce fast, accurate classification systems for 3D MCAD data using significantly smaller datasets than those used for transfer learning in MVCNNs.