基于多编码器和Residual-Transformer的点云补全网络  

Point Cloud Completion Network Based on Multiencoders and Residual-Transformer

作　　者：高辉 杨志景[1] 凌永权 曹江中[1] 李为杰 Gao Hui;Yang Zhijing;Ling Wing-Kuen;Cao Jiangzhong;Li Weijie(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,Guangdong,China)

机构地区：[1]广东工业大学信息工程学院,广东广州510006

出　　处：《激光与光电子学进展》2023年第2期153-160,共8页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(U1701266);广东省自然科学基金(2021A1515011341);广州市科技计划(202002030386)。

摘　　要：点云数据具有无序和稀疏的特点。通过不完整点云数据恢复丢失的三维几何形状的3D点云补全任务是3D视觉技术中一个具有挑战性的问题。现有的3D点云补全网络一般都通过编码器-解码器(Encoder-Decoder)模型直接从部分点云预测完整的点云形状,这会干扰原始部分点云,引入噪声,导致几何位移损失。因此提出一个端到端的网络模型,集中生成平滑和分布均匀的点云对象。所提网络模型主要包含三部分:缺失点云预测、点云融合和点云平滑。第一个模块主要通过多编码器从残缺的点云对象提取局部和全局信息,预测缺失几何部分。第二个模块通过采样算法融合点云。第三个模块基于Residual-Transformer (RT)预测点位移,在避免破坏原始输入点云的空间结构下,可以使点分布得更加均匀。在基准数据集Shapenet-Part上,大量的实验结果表明,所提网络在3D形状补全方面取得了更好的量化结果和更好的视觉效果。Point cloud data has the characteristics of disorder and sparsity.The three-dimensional(3D) point cloud completion task of recovering the missing 3D geometric shapes through incomplete point cloud data is a challenging issue in 3D vision technology.The existing 3D point cloud completion network predicts the complete point cloud shape directly from a subset of the point cloud using the Encoder-Decoder model,which interferes with the original part of the point cloud,resulting in noise and geometric displacement loss.In this study,an end-to-end network model is proposed,which focuses on generating a smooth and uniformly distributed point cloud object.The proposed network model mainly consists of the following three parts:missing point cloud prediction,point cloud fusion,and point cloud smoothing.The first module mainly uses multiencoders to extract local information and global information from incomplete point cloud objects to predict the missing geometric parts.The second module merges point cloud objects by sampling algorithm.The third module is based on a Residual-Transformer(RT) to predict the displacement of the points,which can make the point distribution more uniform without destroying the spatial structure of the original input point cloud.On the benchmark dataset,Shapenet-Part,several experimental results indicate that the proposed network has achieved better quantitative results and visual effects in 3D shape completion.

关 键 词：点云 3D点云补全 自注意力机制 多编码器 残差网络 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]