A deep learning approach for 3D building exterior analysis

Abstract

Advances in the fields of Machine Learning (ML), especially in deep learning, computer vision and computer graphics boosted research in many areas. To this end, and in the best of our knowledge, there has not been any work related with the semantic segmentation of 3D building exteriors, which is the focus of this thesis. In our research, we will employ a volumetric implementation of HRNet and train it for the building semantic segmentation task. For the training purposes of the network we acquired a dataset with 3D models of buildings from Trimble 3DWarehouse and a smaller one with digitisations of monuments located in Cyprus. In our work we try to identify the network setting that gives the best performance, along with the introduction of variations in the input features. Then, having a good backbone with the network trained on the larger dataset, we will fine-tune it to learn the more specific features located in the Cypriot dataset, which because of its size cannot be independently used for the training of a deep network. Finally, with our work we aim to set a stepping stone for future research carried in the field of building analysis.