Feasibility of VR-generated synthetic data for automated productivity monitoring in modular construction

Abstract

This paper examines the feasibility of using VR-generated synthetic data for automated productivity monitoring in modular integrated construction. Site conditions including module shape, color, and occlusion were analyzed to assess their impact on object detection models, and models trained on real world, synthetic, and hybrid datasets were compared. Results showed that the hybrid dataset (real + synthetic) improved detection accuracy, with a 1:3 real to synthetic data ratio yielding the highest performance in this experiment (mean precision = 0.846, recall = 0.88, mAP = 89.7%). While synthetic data enhanced data diversity and detection performance, excessive reliance introduced domain gaps, highlighting the need for a balanced dataset. This paper demonstrates that VR-generated synthetic data can complement real world data, addressing data scarcity in construction site monitoring. The findings contribute to improving AI-driven productivity analysis by optimizing dataset composition and enhancing object detection accuracy in construction automation.