Industrial safety AI needs photos of people near heavy machinery to prevent accidents.
Real-world datasets lack images of the most dangerous scenarios, like workers in blind spots or children near moving equipment.
Collecting this data is unsafe, unethical, and costs between three and five dollars per image for manual annotation.
A new pipeline uses Amazon SageMaker AI and Amazon Rekognition to generate labeled training images automatically.
The system edits real photos of machinery by inserting synthetic people using the Qwen-Image-Edit-2509 diffusion model.
Amazon Rekognition then creates bounding box labels with an 80 percent confidence threshold, converting them to YOLO format.
This method avoids domain gaps and preserves background fidelity while eliminating the need for hazardous photography sessions.
