Research Information System
IEEE International Symposium on Geoscience and Remote Sensing (IGARSS), Athens, Greece, 7 - 12 July 2024, pp.1-4, (Full Text)
Existing supervised point cloud classification methods cannot correctly classify under-represented objects, such as poles, due to the limited number of points in annotated clouds. In this article, we innovatively overcame the challenge of extracting pole-like objects from point clouds proposing a graph-based method. We approach and treat the problem as a hierarchical graph problem, where the focus is on counting the directed edges connecting nodes. The proposed method effectively captures the vertical structure of poles, a key characteristic often missed by conventional deep learning methods due to the sparsity of pole-like objects and the limited number of annotated points. Experiments on the Hessigheim 3D dataset demonstrate the efficiency of the proposed method for automatically extract pole-like objects from airborne point clouds.