Image-based scan-to-BIM for interior building component reconstruction

Abstract

Image-based scan-to-BIM is a cost-effective and accessible solution for generating digital models of real-world environments. However, its indoor application remains challenging due to cluttered occlusions, complex geometries, and various surfaces. This paper develops a photogrammetry and instance segmentation-integrated approach for image-based interior building component reconstruction. Specifically, the approach consists of (1) boundary surface modeling by integrating vertical surface representations and concave polygons, (2) semantic mapping of building components between 3D point clouds and 2D images using learning-based instance segmentation and camera projection, and (3) boundary refinement based on hole and color features for optimizing elements' geometries. The approach is validated using six interior scenes, which shows around 60 % reduction in geometric deviations (56 mm) compared to existing approaches, with mean intersection-over-union ratios of 82 %, 78 %, and 72 % for doors, windows, and lift openings. The approach provides centimeter-level accuracy using commonly available devices, striving to broaden the application of image-based scan-to-BIM.