The 3-D digital representations of building models from Airborne Laser Scan (ALS) point clouds have many
applications in geographic information systems (GIS), remote sensing, archaeology, photogrammetry and,
computer vision. In these applications, especially in urban modelling, there is a strong requirement for
accurate analysis of the building point clouds captured by ALS. Since the roof is the most informative part
of the building in the ALS, 3-D modeling of buildings based on various roof styles is significant, and prior
knowledge about roof style is advantageous for many applications. Automatic roof-top classification,
retrieval, and shape completion using ALS building point clouds are relevant in this context, and
performing these tasks with high accuracy is a great challenge. This is due to the imperfections present in
the ALS roof point clouds such as noise, sparsity, outliers and missing regions. Existing geometric methods
are not suitable for large-sale urban point clouds, complex-shaped buildings as they rely on fixed
geometric priors. A recent trend is to apply machine learning (ML) methods that have better performance,
but classical ML methods use hand-crafted features that reduces its generalization capacity. Hence, there
is a vital requirement for fully automatic methods that are robust to these issues. The advent of deep
learning (DL) has led to significant improvements across various 3-D point cloud analysis tasks, but
standard benchmarks do not evaluate performance on real datasets like ALS point clouds.
Inspired by the success of attention-based methods (like transformers) in Natural Language Processing,
Computer Vision, and Robotics, our study focuses on incorporating attention, introducing fully attentionbased architectures for classification, retrieval, and shape completion tasks of 3-D point clouds. Detailed
experiments suggest the effectiveness of the proposed methods in all these tasks for ALS roof point
clouds. Additionally, we ensure that the models showcase competitive performance in standard
benchmarks and perform detailed robustness tests. 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.