MVCNN++: CAD model shape classification and retrieval using multi-view convolutional neural networks

Deep neural networks have shown promising success towards the classification and retrieval tasks for images and text data. While there have been several implementations of deep networks in the area of computer graphics, these algorithms do not translate easily across different datasets, especially for shapes used in product design and manufacturing domain. Unlike datasets used in the 3D shape classification and retrieval in the computer graphics domain, engineering level description of 3D models do not yield themselves to neat distinct classes. The current study looks at an improved form of the 3D shape deep learning algorithm for classification and retrieval through the use of techniques such as relaxed classification, use of prime angled camera angles for capturing feature detail and transfer learning for reducing the amount of data and processing time needed to train shape recognition algorithms. The proposed algorithm (MVCNN++) builds on top of multi-view convolutional neural network (MVCNN) algorithm, improving its efficacy for manufacturing part classification by enabling use of part metadata, yielding an improvement of almost 6% over the original version. With the explosive growth of 3D product models available in publicly available repositories, search and discovery of relevant models is critical to democratizing access to design models.