Research On 3D Model Classification Based On Graph Convolution

With the continuous development of technology,the application of 3D model is more and more extensive.The task of 3D model classification has gradually become a key task in the industry.With the continuous development of automatic driving,intelligent vehicles,smart cities and other fields,the requirements of 3D model classification tasks are becoming higher and higher.How to classify 3D models efficiently and accurately has become an urgent problem.This paper mainly studies the classification of 3D point cloud data.It mainly includes how to use graph convolution neural network(GCN)to process point cloud data.It also includes how to use 3D shape descriptor as the supplement to local features of point cloud.In addition,GCN is used to extract more comprehensive features,and then Support Vector Machine(SVM)is used to obtain final classification results.For the rotation invariance of point cloud,T-Net is added into GCN.The improved network is used to classify 3D point cloud.The main research contents of this paper are as follows:(1)The method of 3D model classification based on graph convolution is studied.GCN is used to extract features of point cloud.The network structure is designed with reference to existing network models.The aggregation of graph convolution is studied,and Model Net data set is used for classification experiments.Experimental results show that when the aggregation is not too smooth,graph convolution can achieve better effects.(2)The extraction of 3D shape descriptors is studied.D1,D2,D3 and A3 are used to express local geometric features of point clouds.This paper studies the classification method of 3D model that fuses point cloud and 3D shape descriptor.3D shape descriptor is used as local feature representation of point cloud,and point cloud data is fused with 3D shape descriptor.Graph convolution is adopted to extract more comprehensive feature representation.Experimental results show that this method can improve the classification effect.(3)The improved methods of classifying 3D models are studied.One method is that GCN is used to extract features,and then SVM is adopted to process the comprehensive feature representation to improve the classification effect.The other method is to improve the network structure.T-Net is used to learn the rotation invariance of point clouds,so as to the impact of this feature on classification is weakened and the accuracy of classification is improved.Experimental results show that the classification effect of these two methods has been improved to some extent.At the same time,their advantages and disadvantages are introduced.