Cartographic Generalization is a fundamental process in automatic map generation that mostly relies on rule-based constraints. Previous methods that aim at learning this generalization process from data either feed rasterized grids into Convolutional Neural Networks (CNN) or compute hand-designed features to process vector data with Graph Neural Networks (GNN’s). However, there is little work that investigates the application of Deep Learning models directly on the vector coordinates. In this work we demonstrate the efficacy of CNN based architectures to highly constrained geometrical spaces such as building footprints, eliminating the need for more complex architectures. To remedy the lack of annotated data we propose to train geometrical feature embeddings in a self-supervised fashion to directly approximate geometrical properties of local triangles in the building footprints, rather than manually engineering Geometrical Features. We validate our approach on a Building Classification and cartographic generalisation task, outperforming previous methods.