ABSTRACT
This chapter addresses the semantic segmentation of dense point clouds. For this purpose, different strategies may be applied. On the one hand, the rather classic strategy may be applied, which relies on the use of hand-crafted features serving as input to a standard classification technique, which, in turn, delivers a semantic labeling with respect to defined object classes. On the other hand, the strategy relying on the interplay between traditional classification and segmentation techniques allows assigning both class-aware labels with respect to defined object classes and instance-aware labels with respect to objects in the considered scene. Beyond these strategies, the strategy of involving modern deep learning techniques is nowadays commonly applied, as it tends to outperform the other strategies on recent benchmark datasets. After revisiting the fundamentals of the three strategies, a variety of commonly used benchmark datasets for evaluating the performance of a semantic segmentation approach are presented, whereby most of these datasets have been acquired via terrestrial or mobile laser scanning within urban areas.
