FROM MULTI-VIEW TO POINT CLOUD SEGMENTATION: THE CASE STUDY OF VILLA ROBERTI BRUGINE

Point cloud semantic segmentation is a key step for automatically deriving an informative building model from the 3D data reconstruction obtained by 3D surveying tools, such as laser scanners and photogrammetry. Such representation increases the richness of the information available of the represented building, leading to an at least rough interpretation of the scene, and, in particular, a discrimination between the different constitutive elements of the building. The growing interest in semantic building models recently motivated the development of several approaches aiming at obtaining an automatic semantic segmentation of a building point cloud. Such methods are usually either based on the direct segmentation of the point cloud, or on the segmentation of images of the building, then back-projecting the obtained segmentation on the point cloud. Similarly to the latter approach, this work assumes that a proper neural network is available in order to compute the semantic segmentation of building images, and it compares two different strategies for transferring such semantic information from the 2D images to the 3D point cloud. The results obtained in the case study of villa Roberti Brugine (Brugine, Padua, Italy) show that transferring the semantic information can be done quite effectively with the proposed, even when dealing with a certain amount of misclassified points. In particular, best results are obtained in our tests when determining a point class as the most popular classification of such point once projected on all the images where it is visible.