密度导向的点云动态图卷积网络

Density-Oriented Dynamic Graph Convolutional Network of Point Cloud

针对现有主流网络对于点云局部特征提取的能力不足,以及在特征提取过程中未考虑点云密度的问题,提出一种密度导向的点云动态图卷积网络.首先提出点云局部密度指数的概念,衡量点及其邻域点在相应的空间位置中的密集程度;然后利用局部密度指数动态赋予每个点一个膨胀因子,提出密度导向的动态点分组方法对点云构建局部图结构,对每个局部图结构构造动态边缘卷积模块进行特征的提取与聚合,既提取了点云的几何特征,又实现了置换不变性;最后采用残差网络的思想优化图神经网络的过平滑问题.实验结果表明,在分类数据集ModelNet40与ScanObjectNN上,所提网络的分类准确率分别为93.5%和82.2%;在分割数据集ShapeNet与S3DIS上,该网络的平均交并比分别为85.6%和60.4%,均高于DGCNN等主流网络;所提网络在多项任务中的精度都得到显著提升,且在处理密度不均的点云时有较好的鲁棒性,验证了所提算法的可行性与有效性.

A density-oriented point cloud dynamic graph convolutional network is proposed to overcome the shortage in local feature extraction and ignoring the density of point clouds of current methods.Firstly,the point cloud local density index is generated to measure the density of points and their neighborhood points in the corresponding spatial location.Secondly,the local density index is used to dynamically give each point a dilated factor and group points as a local graph.The dynamic edge convolution extracts feature of each local graph.It not only extracts the geometric features,but also realizes the permutation invariance.Finally,the idea of residual network is used to optimize over-smoothing of graph convolutional network.The classifica-tion accuracy of the proposed network on ModelNet40 and ScanObjectNN is 93.5%and 82.2%respectively,and the mean IoU on the segmentation datasets ShapeNet and S3DIS are 85.6%and 60.4%respectively,which are higher than those of current networks such as DGCNN.The experimental results show that the accuracy in these tasks significantly improved,and it has good robustness in processing uneven density point clouds,which verifies the feasibility and effectiveness of the network.