Semantic scene graph-driven indoor image localization in BIM using synthetic views
Journal
Automation in Construction
Journal Volume
187
Start Page
106979
ISSN
09265805
Date Issued
2026-07
Author(s)
Hsu, Wei-Yi
Le, Thai-Hoa
Xiong, Guan-Yong
Chang, Ju-Chi
Lin, Tzu-Yang
Chang, Ting-Wei
Abstract
Accurately localizing indoor inspection images within a Building Information Modelling (BIM) environment is critical for effective facility management and inspection documentation. However, this task remains challenging due to the lack of spatial information in onsite photos and the difficulty of directly linking real images to BIM geometry and semantics. To address these challenges, this paper presents an automated indoor image localization framework including three stages: BIM Mapping, Coarse Localization, and Fine Localization. The framework links real inspection images to BIM by using scene graph representations and a hierarchical localization process. In the BIM Mapping stage, synthetic data, wireframe abstractions, and scene graphs are generated to capture the geometric and semantic structure of indoor spaces. Coarse Localization selects the BIM-rendered view with the most similar scene graph. Fine Localization then refines the camera pose through area-based searching and geometric alignment. Experiments conducted in a research building demonstrate that the proposed method provides reliable indoor localization across different spaces. The approach achieves an average translation error of 1.94 meters and an average rotation error of 34.01 degrees.
Subjects
Building information modeling (BIM)
Camera pose
Image localization
Scene graph
Synthetic
Publisher
Elsevier B.V.
Type
journal article