Large-scale 3D Point Cloud Semantic Segmentation with 3D U-Net ASPP Sparse CNN

Abstract

3D geometric modelling of urban areas has the potential for further development, not only for 3D urban visualization. 3D point cloud, as 3D data commonly used in 3D urban geometry modelling, is needed to extract objects from point clouds to analyze urban landscapes. An automated method to analyze objects from the 3D point cloud can be achieved by using the semantic segmentation method. Unlike other segmentation tasks in 3D point cloud data, 3D urban point cloud segmentation has the challenge of segmenting different object sizes on various types of landscape contours with imbalanced distribution of the object. Therefore, this study modified 3D U-Net Sparse CNN by adding Atrous Spatial Pyramid Pooling (ASPP) as one of the modules in this model, called 3D U-Net ASPP Sparse CNN. The use of ASPP aims to get the contextual multi-scale information of the input feature map from the encoder part of U-Net. Furthermore, 3D U-Net ASPP Sparse CNN is implemented by using weighted dice loss as the loss function. The experiment result shows 3D U-Net ASPP Sparse CNN with weighted dice loss has achieved the best evaluation score in our experiment, with OA = 96.53 and mIoU = 63.59.