Synthetic Data Boosts Industrial Safety AI 160%

Industrial AI safety systems struggle with a critical shortage of training images showing people near dangerous machinery. Staging such scenarios is unsafe, rare in real datasets, costly to annotate, and challenging for edge-deployed models with limited capacity. Researchers built a two-stage pipeline on Amazon SageMaker AI using a diffusion model to insert synthetic people into real equipment images, then automatically labeled them via Amazon Rekognition. The approach achieved up to 160 percent improvement in person detection accuracy without manual annotation or hazardous photography sessions.