Enhancing industrial safety AI with synthetic data on Amazon SageMaker AI
aws ·
Industrial safety AI often lacks training images for rare, dangerous situations involving people near heavy machinery, and staging such scenes for collection is unsafe, unethical, or impractical. The post describes an end-to-end pipeline on Amazon SageMaker AI and Amazon Rekognition that edits real equipment images with Qwen-Image-Edit-2509 to insert realistic people, then generates bounding-box labels automatically. By starting from real scenes rather than creating them entirely synthetically, the approach aims to preserve background fidelity and reduce domain gap while avoiding manual annotation. In experiments using OpenImages train images as a proxy dataset and YOLO11 detectors, placing synthetic people in domain-relevant hazardous positions was more beneficial than placing them in the background, and performance peaked at a tested volume of 750 synthetic images. That setting improved person-detection mAP50 by 160 percent over the real-data-only baseline, while a medium YOLO11 model achieved the best aggregate performance and doubled person recall relative to the baseline. The authors caution that additional synthetic data can introduce generation artifacts and that model capacity should match the available data, while presenting the code and deployment instructions in a GitHub repository.
aws