3D URBAN CHANGE DETECTION WITH POINT CLOUD SIAMESE NETWORKS

Abstract. As the majority of earth population is living in urban environments, cities are continuously evolving and efficient monitoring tools needed to retrieve classify their evolution. In this context, analysing changes between two dates a crucial point. most occur along vertical axis (with new construction or demolition buildings) use 3D data therefore mandatory. Among them, LiDAR constitutes valuable source information. However, With difficulty processing sparse unordered point clouds, existing methods start by rasterizing clouds (for example Digital Surface Models) before using more conventional image tools. This implies significant loss studies dealing directly with best our knowledge, no deep neural network-based method has been explored yet. Thus, order fill gap test ability deal change detection characterization we propose Siamese network Kernel Point Convolution inspired architectures that have already shown performances on 2D images KPConv which achieves high-quality results for semantic segmentation raw clouds. We show quantitatively qualitatively outperforms than 25% (in terms average Intersection over Union classes change) machine learning based hand-crafted features.