Automated Dimension-Aware 2D-To-BIM Reconstruction Through Cross-Modal Text-Geometry Alignment

Abstract: Digitising existing buildings supports energy-efficient retrofitting and lifecycle analysis, yet many are documented only by legacy drawings with dimensions embedded as textual annotations. Existing methods often rely on heuristic global scaling over metric fidelity. This paper presents a vision-based framework integrating optical character recognition (OCR), OCR-guided denoising, semantic segmentation, and grid-based alignment. The primary contribution is a cross-modal text–geometry alignment strategy that reconstructs a metric reference grid from annotations and links it to building geometry. A second contribution is an Industry Foundation Classes-based evaluation framework. Results demonstrate reliable recovery of absolute dimensions for dimension-aware BIM reconstruction.