基于扩张空间金字塔的点云语义分割网络

Point cloud is a sparse, irregular, and unstructured form of data. These properties of the point cloud make it impossible to efficiently obtain the uniform down-sampling of the point cloud, and it becomes difficult to capture the contextual information in the point cloud. High-cost downsampling methods and complex network layers design make existing deep networks can not directly applicable to large point clouds. %The RandLA-Net explored the use of random downsampling for the first time, and achieved a great improvement in efficiency. However, in order to compensate for the information loss caused by random downsampling, the network layer of RandLA-Net method requires a large number of parameters and a well-designed complex encoder layer to obtain semantic wider information, which leads to a large amount of model calculation and limits the scale of point cloud processing on the network.%Existing deep neural networks for point clouds often require a large number of parameters and well-designed complex encoder layers to obtain semantic information, which leads to the computationally heavy model and limits the scale of point clouds processing on the network. In this paper, we explored a new structure for point clouds semantic segmentation deep network. The proposed scalability module uses the idea of dilated convolution in 2D image processing to enlarge receptive fields without downsampling operation, which is computationally expensive for large-scale point clouds.In order to further optimize the performance of the network, we combine the dilated neighborhood with the spatial pyramid structure to replace the complex network layer in the existing methods. Dilated convolution reduces the information loss in the down-sampling operation and simplifies the complex network layers designed to compensate for the loss in the random downsampling. Combining the expanded neighborhood with the spatial pyramid structure to replace the complex network layer in the existing method can significantly reduce network parameters and improve the accuracy rate. %This efficient module can replace several deep complex decoder layers in a deep network for semantic information extraction, reducing network depth, which in turn greatly reduces network size and consumption. The proposed model DSPNet, with RandLA-Net as a backbone network, achieves more superior accuracy and efficiency. We tested accuracy on SemanticKITTI,and Semantic3D datasets. We also made experiments to compare our model with other state-of-the-art methods. The simulation results shows that compared to the baseline model, our method reduces the number of parameters by 59\% with similar accuracy, and has faster time and more inference points.