Point Cloud Shape Completion Method Based On Attention And Dynamic Graph Mechanism

With the popularity of 3D sensors,point cloud has been widely used in various fields of 3D vision.Due to the limitation of occlusion,light reflection,transparency of material surface,resolution,and viewing angle of sensor,point cloud will be missing and sparse.Therefore,point cloud shape completion is an important research topic.In recent years,some progress has been made in using deep learning technology to complete the shape of the point cloud,but obtaining accurate complete shape descriptions from the incomplete point cloud is still a challenging problem.This dissertation mainly studies a point cloud completion method based on attention and dynamic graph mechanism and deeply studies the way of point cloud detail feature extraction and the way of capturing the structural relationship between points.Based on the existing methods,a series of efficient network architectures are designed to improve the performance of the point cloud shape completion tasks.The main work of this dissertation is as follows:1.A point cloud shape completion network based on generative adversarial network and self-attention is constructed.The architecture consists of a self-attention module and a generative adversarial network.First,use multilayer perceptron to effectively extract the global features of the point cloud,and add a self-attention module to capture local details.Then,the feature pyramid module is used to complete the generation of main center point coordinates,secondary center point coordinates,and fine node coordinates.Finally,the generated point cloud is optimized by using the discriminator with the self-attention module,to make the results more accurate,the network can effectively extract the characteristics of the point cloud and complete the missing part of point cloud.And optimize the generated point cloud in a confrontational way.A large number of experiments show that on Shape Net dataset,has excellent completion performance.2.A point cloud shape completion network based on the 3D grid method is designed.The network extracts the features of incomplete point clouds from coarse to fine,including Gridding of the point cloud,feature sampling and transformer module.Firstly,in the point cloud processing stage,the point cloud is mapped into mesh to obtain more sufficient feature representation.Then,a global feature encoder is designed based on transformer module to better extract point cloud features.Finally,the advanced feature representation of point cloud is gradually focused on channel,local space and non-local space,and the feature representation of point cloud is optimized and adjusted to generate a completion point cloud.Experiments on Shape Net and KITTI datasets show that the proposed network can effectively improve the performance of point cloud completion.3.3D point cloud shape completion network based on dynamic graph convolution is constructed.Firstly,drawing on the idea of graph convolution,edge convolution is used to extract the local neighborhood features of the point cloud,the local geometric features are captured while maintaining the invariance of the arrangement.Learning the best data enhancement method of the current task to further improve the complete performance.Next,the global feature representation is obtained by global max pooling and average pooling.After that,a feature pyramid network is designed to gradually generate a point cloud to complete the missing point cloud.Last,experiments verify the effectiveness of the network in point cloud completion.