基于Spider卷积的三维点云分类与分割网络

针对传统的卷积神经网络(CNN)不能直接处理点云数据,需先将点云数据转换为多视图或者体素化网格,导致过程复杂且点云识别精度低的问题,提出一种新型的点云分类与分割网络Linked-Spider CNN。首先,在Spider CNN基础上通过增加Spider卷积层数以获取点云深层次特征;其次,引入残差网络的思想在每层Spider卷积增加短连接构成残差块;然后,将每层残差块的输出特征进行拼接融合形成点云特征;最后,使用三层全连接层对点云特征进行分类或者利用多层卷积层对点云特征进行分割。在ModelNet40和ShapeNet Parts数据集上将所提网络与PointNet、PointNet++和Spider CNN等网络进行对比实验,实验结果表明,所提网络可以提高点云的分类精度和分割效果,说明该网络具有更快的收敛速度和更强的鲁棒性。

3D point cloud classification and segmentation network based on Spider convolution

The traditional Convolutional Neural Network (CNN) cannot directly process point cloud data, and the point cloud data must be converted into a multi-view or voxelized grid, which leads to a complicated process and low point cloud recognition accuracy. Aiming at the problem, a new point cloud classification and segmentation network called Linked-Spider CNN was proposed. Firstly, the deep features of point cloud were extracted by adding more Spider convolution layers based on Spider CNN. Secondly, by introducing the idea of residual network, short links were added to every Spider convolution layer to form residual blocks. Thirdly, the output features of each layer of residual blocks were spliced and fused to form the point cloud features. Finally, the point cloud features were classified by three-layer fully connected layers or segmented by multiple convolution layers. The proposed network was compared with other networks such as PointNet, PointNet++ and Spider CNN on ModelNet40 and ShapeNet Parts datasets. The experimental results show that the proposed network can improve the classification accuracy and segmentation effect of point clouds, and it has faster convergence speed and stronger robustness.