Towards accurate urban scene understanding using point clouds: The SemanticUrban dataset

Point clouds are essential data representations of three-dimensional surfaces of real-world scenes and objects. With the recent developments in urban scene understanding, there is a substantial demand for semantic point cloud datasets that represent urban scenes with high semantic accuracy and fine semantic details. However, existing benchmark datasets of this kind are very limited, and their semantic information is either not accurate enough or lacks detailed semantic classification. In this paper, we present SemanticUrban, a large-scale high-resolution point cloud dataset acquired from 150 urban scenes using terrestrial laser scanning. SemanticUrban features super-high-resolution point clouds and highly accurate semantic categorizations, classifying each data point into one of 23 defined classes. We conduct extensive evaluations on SemanticUrban using representative deep learning methods, followed by a detailed discussion of our findings. Additionally, we highlight main challenges associated with the SemanticUrban dataset, motivating future research to develop new approaches for tackling these issues. The SemanticUrban dataset is available at https://github.com/YuanFangFF/SemanticUrban.