Having access to accurate and recent digital twins of infrastructure assets benefits the renovation, maintenance, condition monitoring, and construction planning of infrastructural projects. There are many cases where such a digital twin does not yet exist, such as for legacy structures. In order to create such a digital twin, a mobile laser scanner can be used to capture the geometric representation of the structure. With the aid of semantic segmentation, the scene can be decomposed into different object classes. This decomposition can then be used to retrieve cad models from a cad library to create an accurate digital twin. This study explores three deep-learning-based models for semantic segmentation of point clouds in a practical real-world setting: PointNet++, SuperPoint Graph, and Point Transformer. This study focuses on the use case of catenary arches of the Dutch railway system in collaboration with Strukton Rail, a major contractor for rail projects. A challenging, varied, high-resolution, and annotated dataset for evaluating point cloud segmentation models in railway settings is presented. The dataset contains 14 individually labelled classes and is the first of its kind to be made publicly available. A modified PointNet++ model achieved the best mean class Intersection over Union (IoU) of 71% for the semantic segmentation task on this new, diverse, and challenging dataset.