MVCNN plus plus : Computer-Aided Design Model Shape Classification and Retrieval Using Multi-View Convolutional Neural Networks

Deep neural networks (DNNs) have been successful in classification and retrieval tasks of images and text, as well as in the graphics domain. However, these DNNs algorithms do not translate to 3D engineering models used in the product design and manufacturing. This paper studies the use of multi-view convolutional neural network (MVCNN) algorithm enhanced by the addition of engineering metadata, for classification and retrieval of 3D computer-aided design (CAD) models. The proposed algorithm (MVCNN++) builds on the MVCNN algorithm with the addition of part dimension data, improving its efficacy for manufacturing part classification and yielding an improvement in classification accuracy of 5.8% over the original version. Unlike datasets used for 3D shape classification and retrieval in the computer graphics domain, engineering level description of 3D CAD models do not yield themselves to neat, distinct classes. Techniques such as relaxed-classification and prime angled cameras for capturing feature detail were used to address training data capture issues specific to 3D CAD models, along with the use of transfer learning to reduce training time. Our study has shown that DNNs can be used to search and discover relevant 3D engineering models in large public repositories, making 3D models accessible to the community.