The 3-D digital representations of building models from Air-borne Laser Scan (ALS) point clouds have many applications in urban modelling, archaeology, photogrammetry, etc and there is a strong requirement for accurate analysis of these scanned 3D point clouds. The existing works involving traditional geometric methods or machine learning are not efficient due to the various imperfections in the ALS data and other
reasons. The advent of deep learning has led to significant improvements across various 3-D point cloud processing tasks but its effectiveness is not yet fully explored in remote sensing and GIS. In this study, we introduce novel deep learning methods for shape classification, retrieval and completion of ALS roof point clouds.
Inspired by the success of attention-based methods (like transformers) in Natural Language Processing, our study focuses on incorporating attention, introducing full-blown attention based architectures for the above tasks. Detailed experiments suggest the effectiveness of the proposed methods in all these tasks. Additionally, we also perform
robustness tests on the real unaligned datasets containing various imperfections. The proposed architectures are also highly time and memory-efficient when compared to other methods. These results point out that attention-based methods might be an even-more natural fit for point cloud processing, as attention in its core is a set operation: implying that it is invariant to permutation and cardinality of the input elements making it ideal for point clouds.