Three-dimensional model retrieval in single category geometry using local ontology created by object part segmentation through deep neural network

3D model retrieval is useful for reusing designs in manufacturing industry. Traditionally, 3D model retrieval has been implemented only in low-level information such as geometry, color, and texture. However high-level semantic information should be used for more accurate retrieval. In this study, a 3D geometry is divided into several parts using PointNet and then the local ontology is constructed by summarizing the characteristics of each part. Then part align similarity, lemma similarity, name similarity, part location similarity, and part size similarity are calculated. Using the values of these similarities, 3D models are retrieved from input query model. This comprehensive retrieval that includes all the similarities is more balanced and shows better performance in nameless models than considering only partial similarities. Through the method in this paper, high-level information and low-level information can be used simultaneously for 3D model retrieval.