| In recent years,deep neural network has achieved remarkable results in many tra-ditional cognitive tasks such as speech recognition,target detection and machine trans-lation.Deep learning is more powerful,especially in the analysis of 2D image analysis.Due to the success of deep learning in 2D images and the blowout of 3D image data,the deep neural network has been applied to the analysis of point clouds.The main tasks of large-scale point cloud image analysis focus on semantic segmentation and instance segmentation,especially for the problem of point cloud instance segmentation,which is difficult to be completed in high speed,high efficiency and high accuracy.At present,the semantic segmentation network has become relatively mature due to the PointNet network in the area of point cloud segmentation.The current mainstream point cloud instance segmentation methods are mainly divided into two steps.Firstly,the points of each instance are aggregated in the feature space by means of measurement learning to make the features of the same instance as similar as possible.Then,the ag-gregated vector cluster is segmented to build a proposal for each instance.Much of the previous work focused on the aggregation of the more basic vectors,with a simple treat-ment of how the aggregated instances are divided.In order to solve the above problems,the point cloud segmentation algorithm based on seed point selection is proposed in this paper.Based on the judgment of "seedness" of each point,the seed selection network selects a better seed point to generate a proposal,and then completes the instance-level segmentation of all points.In addition,the speed of instance segmentation is effectively improved by rapidly processing the points and low "seedness" points of the generated instances.In the experiment,the selection network of seed points is grafted on different instance segmentation networks,and the accuracy and efficiency of segmentation are improved in different degrees. |
PDF Full Text Request