Augmenting Construction Safety Datasets with Training-Free Diffusion-Based PPE Editing

Deep learning-based computer vision plays a critical role in monitoring construction site safety. However, its performance is often constrained by highly imbalanced datasets, particularly for personal protective equipment (PPE) such as helmets, vests, and safety shoes. To address this limitation, we present a learning-free conditional augmentation framework built on pre-trained diffusion models. Our method integrates (i) automatic PPE region segmentation using Grounded-SAM, and (ii) diffusion-based inpainting guided by positive/negative prompts to create controlled variations of real construction site images. Unlike conventional generative or simulation-based approaches, our pipeline requires no model retraining, preserves the realism of original scenes, and allows fine-grained attribute toggling (e.g., removing helmets, switching footwear, or altering uniforms). Experimental results on 1,000 images demonstrate that our approach achieves substantially lower FID scores compared to the baseline only with prompts. This framework provides a cost-effective and reproducible solution for generating safety-critical visual data, thereby facilitating the broader adoption of AI-based monitoring systems in real-world construction environments.