基于动态图卷积网络的点云分类和分割网络

Point Cloud Classification and Segmentation Network Based on Dynamic Graph Convolutional Network

点云分类和分割是三维场景理解中的关键步骤。针对稀疏点云输入和遮挡不能有效识别点云的问题,提出一种改进型分类和分割网络Linked-DGCNN。在动态图卷积网络(DGCNN)的基础上增加EdgeConv卷积层数以提取深层次点云特征;去除DGCNN的转换网络以简化网络结构;引入深度残差网络的思想连接不同网络层的输出特征,形成点云特征,同时使网络训练更加稳定。基于ModelNet40和ShapeNet Parts数据集将该网络与其他点云网络进行对比实验,实验结果表明,该网络在稀疏点云输入和遮挡情况下,相比其他方法有较高的点云分类和分割精度,由此说明该网络具有较强的鲁棒性。

Point cloud classification and segmentation are key steps in understanding three-dimensional(3D)scenes.Aiming at the problem that sparse point cloud input and occlusion cannot effectively identify point clouds,an improved classification and segmentation network linked-dynamic graph convolutional neural network(DGCNN)is proposed.First,the deep-level point cloud features were extracted by increasing the number of EdgeConv convolutional layers based on DGCNN.Next,the transformation networks of DGCNN were removed to simplify the network structure.Finally,the idea of introducing a deep residual network was used to link the output features of different network layers to form point cloud features,making the network training more stable.The proposed network was compared with other point cloud networks on ModelNet40 and ShapeNet Parts datasets.The experimental results show that the network has higher accuracy of point cloud classification and segmentation than other methods under the sparse point cloud input and occlusion.Besides,it has stronger robustness.