Deep Neural Networks to Assist in BIM Creation Using Scanned Data: A Review

Abstract

Deep neural networks (DNNs) have been revolutionizing various 3D processing fields, such as autonomous driving and augmented reality. They also have a great potential to facilitate building information modeling (BIM) based on 3D scanned data, a process known as Scan-to-BIM. Several studies have investigated the potential of DNNs to improve the Scan-to-BIM pipeline. However, most of the existing studies are high-level overviews of DNN applications. In this study, an in-depth investigation of the DNNs’ capacities in improving 3D reconstruction, object detection, and object parametrization, which serve as the core components of Scan-to-BIM, is provided. Our investigation is performed by reviewing the most relevant studies of the computer vision and construction literatures to gain a comprehensive view of both the stateof-the-art processing methods of scanned data, as well as the progress of automated BIModeling using Industry Foundation Class (IFC) objects. Based on the reviewed state-of-the-art studies, current challenges and limitations are discussed to identify further avenues for research.