Research On Classification Algorithms Of 3D Point Cloud Object Based On Deep Learning

This thesis presents an end-to-end framework,SK-Net,to jointly optimize the unsupervised inference of spatial keypoints with the learning of feature representation of a point cloud’s local detail and ditribution pattern for a specific 3D task.This framework is used for the point cloud recognition tasks.The research of this topic is based on the National Natural Science Foundation of China" Measure preserving transformation for 3D deformable objects and recogniton of them by sparse manifold embedding methods".The key contributions of this paper are as follows:·SK-Net employs a modified PointNet framework to end-to-end discover spatial keypoints of a point cloud and achieves the spatial modeling of the point cloud.The generation of spatial keypoints is jointly conducted by the proporsed regulating losses and a task objective function without knowledge of spatial keypoints of keypoint location annotations and proposals.Meanwhile,these learned spatial keypoints corresponding to geometrically and semantically meaningful regions of the point cloud.·This paper verifies that the local detail features extracted in the local area of the point cloud corresponding to the learned spatial keypoints are robust to the variability of points density.In addition,various local region sampling approaches are applied for capturing local information.At the point cloud’s distribution space,GCN,LSTM and othe algorighms are used to perform correlation construction and feature coding between different regions.These ablation experiments contribute to determine the best model of the proposed framework.·A extendable module,Reference module,is designed to achieve the instance retriveal task of 3D model.This module can be extended to other 3D classification algorithms,such as PointNet,PointNet++,MVCNN,etc.and greatly improves the algorithms’ retrieval performance on the ModelNet dataset.·In 3D tasks,such as classification and segmentation,the proposed method performs better than or comparable with the state-of-the-art approaches(ModelNet 10/40:96.2%/92.7%accuracy;ShapeNetPart:85,0%mIoU;ScanNet:81.4%per-point classification performance).After adding the Reference module,SK-Net achieves 92.4%mAP in the retrieval task on ModelNet40 and 95.0%mAP on ModelNet10.For inference,13.4 ms is consumed in the classification task.Part of the content of the research results is organized into a paper,which is accepted by CCF-A conference AAAI as an oral paper.