| Point cloud completion has become a popular research area in 3D computer vision,aiming to recover the complete point cloud from its partial observation.Point clouds have been widely researched and applied in various areas because they are one of the most convenient 3D geometric data that people can obtain from the real world.Unfortunately,the point clouds obtained by people tend to be incomplete,with various forms of partiality and holes,which affect downstream tasks.Therefore,people leverage the strong association and prediction ability of neural networks in deep learning to perform point cloud completion.However,previous methods for point cloud completion either directly predicted the whole shape,changed the original distribution of points even resulting in a totally different shape,or had limited performance in reconstructing tiny and detailed object components,making it impossible to utilize geometric characteristics such as symmetry or similarity.In this work,we propose a novel patch-based Dual-Path Network for point cloud completion,increasing the accuracy and visual quality of results while keeping the original distribution of points in the input point cloud.This network recovers the local details of objects more finely by leveraging the advantages of different encoder architectures.The network is divided into two paths,with the upper path providing a rough estimation for the global shape of the missing area,and the lower path filling in the details by generating several point cloud patches based on the results of the other path.We also propose an identifier module to retain the original points in the partial point cloud as much as possible.Comprehensive quantitative and visual experiments were conducted on multiple datasets in this work.Ablation studies were carried out to verify the effectiveness of each module’s design,and robustness tests demonstrated stability against various scales of partiality.These experiments showed the superior performance of our method in accurately completing 3D point clouds while highlighting the usefulness of each module.Our method achieved better numerical and visual results than previous works,even when faced with different missing scales of the point cloud. |
