ControlNet-based domain adaptation for synthetic construction images via graphical simulation and generative AI
Sina Davari, Daeho Kim, Ali Tohidifar · Automation in Construction · 2025
Data scarcity in construction hinders deep neural network training for computer vision applications. While synthetic data generators provide annotated images, they lack realism, creating a reality gap that leads to suboptimal real-world performance. This paper introduces BCon, a framework that integrates BlendCon, a construction data generation engine, with ControlNet, a generative architecture featuring conditioning controls, to enhance the realism and diversity of synthetic images while preserving annotations. Through hyperparameter tuning and post-processing, a dataset of 25,600 enhanced images is created. Quantitative evaluations demonstrate significant improvements in realism metrics: DreamSim (+9.5%), VIEScore (+114.3%), CLIPScore (+14.7%), and FID-5k (+22.6%), indicating closer alignment with real images. Moreover, YOLOv10 models trained on enhanced images achieve an AP 50–95 of 0.66 on worker detection, outperforming those trained on original synthetic data by 7.9% and slightly surpassing models trained on equivalently sized real data. This framework offers cost-effective, high-quality dataset generation for visual AI applications in construction.