3D Point Cloud Analysis Method Based On Point Cloud Feature Space

Deep learning has been widely used in computer graphics in recent years.However,current methods on 3D data based on deep learning,especially the representation,classification,and reconstruction of the point cloud,are still in the initial stage.Current methods have proposed local neighborhood graph convolution based on deep learning to extract local features of point clouds.However,they mainly focus on 3D spatial information,and there are still challenges in extracting geometric information of the surface formed by the point cloud.Although researchers have established public point cloud data sets,the number of 3D models is limited,and it will consume a lot of time and energy to increase 3D data.The size of the training set generally limits current methods.We creatively propose a novel surface experience graph convolution neural network framework SEGCNN to analyze point clouds based on point cloud feature space.SEGCNN consists of two crucial modules,the surface feature graph convolution operation SGConv and the point cloud prior knowledge generated by GU-GAN.SGConv is used to extract spatial relations among 3D points and the variation trend information of the local surface formed by the point cloud.GU-GAN is an unsupervised generative adversarial network framework based on graph convolution,used for warm starting SEGCNN.We pre-train GU-GAN to obtain the point cloud spatial feature vector and send it into SEGCNN as prior knowledge,enabling SEGCNN to quickly perceive spatial relations and the variation trend information of the local surface in fewshot learning.We validate the classification performance of SEGCNN on ModelNet10,and the overall accuracy achieves 95.9%.We also perform the few-shot learning experiments on ModelNet40.When the training set size is reduced to 30%,the overall classification accuracy can reach 91.8%,which is 2.5% higher than Geo-CNN.Experiments demonstrate that our method improves accuracy in point cloud classification tasks and few-shot learning significantly compared to existing approaches such as PointNet,PointNet++,DGCNN,and Geo-CNN.