What's the Point? Using Extended Feature Sets For Semantic Segmentation in Point Clouds

A recent focus on expanding deep learning to use non-traditional input data has seen a high growth in research of deep learning on point sets. Due to its high collection cost and lack of available labeled data, there is an absence of research into deep learning with aerial LiDAR. In this paper, we present a new benchmark labeled dataset, called "Surrey Aerial 3", for evaluating networks on aerial LiDAR data. The dataset covers over 6km(2) and has three classes in multiple environments. We provide our architecture, "Curvature Weighted PointNet++" that eliminates PointNet++'s random batch selection and provides a way to select batches based on key points of interest selected from the Eigen feature space. We extend the hierarchical feature space to add additional layers of context to address the need for an extended field of view in aerial LiDAR.