SEMANTIC SEGMENTATION OF TERRESTRIAL LIDAR DATA USING CO-REGISTERED RGB DATA

Abstract. This paper proposes a semantic segmentation pipeline for terrestrial laser scanning data. We achieve this by combining co-registered RGB and 3D point cloud information. Semantic is performed applying pre-trained off-the-shelf 2D convolutional neural network over set of projected images extracted from panoramic photograph. allows the to exploit visual image features that are learnt in state-of-the-art models trained on very large datasets. The study focuses adoption spherical information capture assessing results using classification metrics. obtained demonstrate approach promising alternative asset identification comparable performance with machine learning frameworks, however, avoid both labelling training efforts required such approaches.