The scarcity of task-aligned training data remains a primary bottleneck in advancing vision-based construction automation. This paper introduces Synthetic Deep Active Learning (SDAL), a fully automated, end-to-end framework that integrates controllable image synthesis, automatic labeling, and active learning in an iterative loop. SDAL employs an Oracle module to automatically identify high-loss cases, generates and labels look-alike synthetic images for these scenarios via a BlendCon module, and retrains the model with increasingly targeted samples. This automated loop eliminates manual labeling, reduces computational burden, and continually improves domain alignment. In construction worker detection, SDAL achieves a + 13.14% AP (0.5:0.95) improvement over benchmark models and outperforms state-of-the-art baselines under equal computational budgets. By enabling models to be rapidly adapted and redeployed across diverse projects with minimal supervision, SDAL introduces another pathway for synthetic data–driven learning in construction, bridging the persistent gap between research innovations and real-world deployment.
Tohidifar et al. (Sun,) studied this question.