Semantic Segmentation of AHN3 Point Clouds with DGCNN

Abstract

Semantic segmentation of aerial point clouds with high accuracy is significant for many geographical applications, but is not trivial since the data is massive and unstructured. In the past few years, deep learning approaches designed for 3D point cloud data have made great progress. Pointwise neural networks, such as PointNet and its extensions, show their ability to process 3D point clouds, especially in classification and semantic segmentation. In this work, we implement DGCNN (Dynamic Graph CNN), which combines PointNet with Graph CNN, and extend its semantic segmentation application from indoor scenes to an aerial point cloud dataset: The Current Elevation File Netherlands (AHN), which was produced by airborne laser scanners for the whole Netherlands. Point clouds from the iteration AHN3 are classified into four classes: ground, building, water and others (including vegetation, railways, etc). Moreover, DGCNN splits the input point cloud into regular blocks before operating on it and processes each block independently, which limits the effective range (receptive field) of the network to some extent. Thus, the second aim of this work is to investigate the impact of the effective range on the performance of DGCNN by adjusting two crucial parameters: the block size and the neighborhood size k in k-NN graphs. It turns out that enlarging the block size or k helps to improve the overall accuracy of DGCNN, but cannot ensure better segmentation results from each individual class. With the block size 50 m and k=20, the most balanced F1 scores for all classes and an overall accuracy of 93.28% are achieved. Based on the evaluation for each setting with a certain block size and k, we also manage to further improve the overall accuracy to 93.51% by combining smaller-scale (with block size 30 m) and larger-scale (with block size 50 m) segmentation results, with k=20.