A systematic review of artificial intelligence for capturing real-world structures into building information modelling

Abstract

The architecture field faces increasing pressure to digitize complex real-world structures, yet traditional Scan-to-Building Information Modelling (Sc2BIM) workflows remain time-consuming and require many resources. This research addresses the challenge of how Artificial Intelligence (AI) can enhance the Sc2BIM process by automating critical tasks such as point cloud segmentation and model generation, which are essential for producing accurate and efficient Building Information Models (BIM). To tackle this problem, we conducted a systematic review following a structured four-stage methodology: we formulated research questions to define the study's scope, applied clear inclusion and exclusion criteria to identify and screen relevant papers from main scientific resources, performed a detailed statistical analysis of selected studies, and synthesized the results to highlight current trends and research gaps. Our findings indicate that PointNet++ is the most frequently used model for 3D point cloud segmentation, while Convolutional Neural Networks (CNNs) remain the dominant architecture overall; hybrid and transformer-based models are getting popular, and some studies demonstrate successful full 3D BIM reconstructions from raw scans. These results underline the growing role of AI in streamlining the Sc2BIM pipeline, potentially reducing manual effort and improving model accuracy. The key research implication is a comprehensive overview of how AI techniques are applied in Sc2BIM, providing valuable insights for researchers and practitioners seeking to advance automation and efficiency in the digitization of the built environment.