Deep generative modelling for building facade footprint extraction from ALS point clouds

Extracting building footprints from airborne laser scanning (ALS) data has been an active topic in remote sensing. The existing methods usually obtain footprints by roof detection, which inevitably leads to inconsistent representation between the extracted and the facade footprints if the roof overhang is present. Meanwhile, the incompleteness of building walls in ALS data makes facade footprint extraction a challenging task. To address these issues, we utilize deep generative modelling to predict building facade footprints directly, while the wall information is implicitly obtained by the neural network. Besides, a contrastive generation strategy is presented to further promote extraction performance. We also build a new challenging ALS dataset, on which the proposed contrastive generation network (CG-Net) achieves 10.13% and 6.36% improvements in intersection over union (IoU) and F1-score, respectively, compared with a typical roof-detection-based method on the building footprint extraction task. The dataset in this study will be made publicly available.