Automated training data creation for semantic segmentation of 3d point clouds

Abstract. The creation of as-built Building Information Modelling (BIM) models currently is mostly manual which makes it time consuming and error prone. A crucial step that remains to be automated is the interpretation of the point clouds and the modelling of the BIM geometry. Research has shown that despite the advancements in semantic segmentation, the Deep Learning (DL) networks that are used in the interpretation do not achieve the necessary accuracy for market adoption. One of the main reasons is a lack of sufficient and representative labelled data to train these models. In this work, the possibility to use already conducted Scan-to-BIM projects to automatically generate highly needed training data in the form of labelled point clouds is investigated. More specifically, a pipeline is presented that uses real-world point clouds and their corresponding manually created BIM models. In doing so, realistic and representative training data is created. The presented paper is focussed on the semantic segmentation of 6 common structure BIM classes, representing the main structure of a building. The experiments show that the pipeline successfully creates new training data for a recent DL network.