Domain adaptation methodology using synthetic images for computer vision-based automated construction progress monitoring

Abstract

Purpose This study aims to improve the generalization of computer vision models used for automated construction progress monitoring by addressing the domain shift problem between synthetic and real images. The objective is to enable more reliable deployment of synthetic data-driven models in actual construction environments. Design/methodology/approach A hybrid optimization algorithm is proposed to minimize domain discrepancy, measured by maximum mean discrepancy (MMD), between synthetic and real images. The method uses a YOLOv8-based feature extractor to compute class-specific distributions and applies both gradient-based and Bayesian optimization techniques to tune image transformation parameters. These transformations are selectively applied to synthetic images to improve alignment with real-world feature distributions. Findings Experimental results show that the proposed method effectively reduces MMD values for both concrete and insulation classes, enhancing visual and structural similarity to real images. Moreover, the use of class-wise single-transformation optimization demonstrated superior performance and stability compared to full-parameter optimization. The hybrid data set combining raw and transformed synthetic images yielded the highest detection accuracy across all test scenarios. Originality/value This study introduces a domain adaptation framework based on MMD optimization specifically tailored for the construction industry. Unlike conventional approaches, the proposed method provides a class-specific, quantitatively driven adaptation strategy that enhances model reliability without requiring extensive real-image data sets. The framework offers practical applicability for progress inference and forms a foundational module for future integration with BIM-based monitoring systems.