3D Shape Representation Learning And Object Detection

Learning the shape information of point clouds is designed to understand the semantic level of point clouds based on the characteristics of the 3d point cloud learned.The current deep learning frameworks rarely take into account the local neighborhood information of the points,which leads to the loss of a part of the local information and the shape characterization of point clouds.In addition,the processing of the invariance of the point cloud and the frame design of the local features cannot obtain the shape information well,which makes the recognition ability stagnation.When designing deep learning frameworks to learn point cloud features,the size of the parameters of the framework is often not considered.In response to the above issues,this article will conduct the following research:First,study the effect of different layers in a 3D point cloud deep learning framework and how the framework can better capture shape information.On the one hand,research on the theory of point cloud shape representation based on deep learning,the role of each part in the framework design is analyzed and the local feature acquisition module in the network framework is redesigned.On the other hand,a multi-scale input order learning module is proposed to better understand the characteristics of the point cloud based on the characteristics of the point cloud.Secondly,the related theories of predicting the center of the object and the bounding box in 3D target detection by studying the local feature learning method which is enhanced by considering the edge features of the neighborhood are studied.According to the two-dimensional mature target detection,the search range of the three-dimensional target is reduced,and the point cloud containing each target is extracted by considering edge features to extract local features to instance segment the object more accurately.Further predicting the center and size of the three-dimensional target can improve the accuracy of 3D object detection,and the average accuracy of the three categories in the KITTI dataset is improved by nearly 1%.Thirdly,Investigate the application of attention mechanism in 3D deep learning and related theories of application methods.The in-depth study is made on how to obtain the local neighborhood information and the attention mechanism in the 3D point cloud.It can enhance local feature learning ability by application the attention mechanism on local neighborhood learning process.The local feature learning module with the attention mechanism can pay more attention to the influence of different points in the local neighborhood on the local structure,and demonstrate the superiority of this method by verification on the ModelNet10,ModelNet40 and ShapeNet dataset,and achieve 93.8%overall accuracy,92.2%overall accuracy,and 84.2%average category accuracy,respectively.