A NEW DATASET AND METHODOLOGY FOR URBAN-SCALE 3D POINT CLOUD CLASSIFICATION

Abstract

Urban landscapes are characterized by a multitude of diverse objects, each bearing unique significance in urban management and development. With the rapid evolution and deployment of Unmanned Aerial Vehicle (UAV) technologies, the 3D surveying of urban areas through high resolution point clouds and orthoimages has become more feasible. This technological leap enhances our capacity to comprehensively capture and analyze urban spaces. This contribution introduces a new urban dataset, called YTU3D, which covers an area of approximately 2 km2 and encompasses 45 distinct classes. Notably, YTU3D exceeds the class diversity of existing datasets, thereby enhancing its suitability for detailed urban analysis tasks. The paper presents also the application of three popular deep learning methods in the context of 3D semantic segmentation, along with a multi-level multi-resolution (MLMR) integration. Significantly, our work marks the first application of deep learning with MLMR in the literature and shows that a MLMR approach can improve the classification accuracy. The YTU3D dataset and research findings are publicly available at https://github.com/3DOM-FBK/YTU3D.