Construction Scene Segmentation Using 3D Point Clouds: A Dataset and Challenges

With the purpose of facilitating process tracking such as inspection reports and progress monitoring, the AEC industry has adopted an as-built 3D model that is reconstructed using a 3D scanner during or after construction. In response to the laborious difficulties of converting a point cloud into a semantically rich model, for example, BIM, researchers are attempting to automate this process via machine learning, applying 3D semantic segmentation and parametric modeling. However, there are no publicly accessible 3D datasets that target construction sites, regarded as unstructured and cluttered scenes, thus yielding a barrier to construction scene segmentation development. To this end, this paper aims to generate a 3D construction dataset that can be utilized for machine learning models requiring ground truth and to suggest foundation processing for general scene segmentation on construction datasets. In addition, we identify and discuss several challenges pertaining to construction sites, in terms of 3D semantic segmentation.