An Efficient 3D Model Retrieval Method Based on Adaptive Clustering

In the field of mechanical design, most designs are achieved by making improvements based on previous design results. Therefore, work efficiency can be improved by retrieving 3D models and selecting similar parts from the database. In view of the situation that the existing 3D model retrieval methods are insufficient in distinguishing intra-class variations, an efficient method for extracting 3D model features is proposed. The output of network combined with geometric feature vectors are clustered by universal ordering points to identify the clustering structure(UOPTICS) algorithm, and the network is trained again using the combined mean square error(CMSE) loss function, so that the network has the ability to distinguish the subcategory when only the parent category labels are used as the training input. It can preferentially retrieve models that are more similar in appearance. On modelnet40 single-modal point cloud data, the mAP reaches 92.8%. In the subcategory-oriented retrieval, the mAP reaches 94.3%.