Describing Paris: Automated 3D Scene Analysis via Distinctive Low-Level Geometric Features

The automated analysis of 3D point clouds has become a topic of great importance in photogrammetry, remote sensing, computer vision and robotics. One avenue of research directly addresses the analysis of urban environments, where recent investigations focus on 3D reconstruction [18, 33, 13], consolidation of imperfect scan data [32, 5], object detection [24, 28, 4], extraction of roads and curbstones or road markings [2, 34, 9], urban accessibility analysis [25], recognition of powerline objects [12], extraction of building structures [27], vegetation mapping [31], large-scale city modeling [14], semantic perception for ground robotics [10] and semantization of complex 3D scenes [1]. A common task for many of these different applications consists of point cloud classification [11, 19], where each 3D point is assigned a specific class label. Addressing the issue of urban point cloud classification – where the spatial 3D data may be collected via airborne, terrestrial and/or mobile laser scanning – we face a variety of challenges arising from the complexity of respective 3D scenes caused by an irregular sampling and very different types of objects. Since the results of urban 3D scene analysis may vary from one dataset to another, publicly available standard datasets are desirable in order to compare the performance of different methodologies. Consequently, there has been a steadily increasing avail-