AWS reports synthetic-data gains for industrial safety detection

AWS reported that a synthetic-data pipeline improved person detection in tests designed around hazardous proximity to industrial equipment. The workflow used a diffusion model hosted on SageMaker AI to insert people into real equipment images, then used Amazon Rekognition to create bounding-box labels. In an OpenImages train-data proxy, AWS said 750 synthetic images raised person mAP50 from 0.051 to 0.134, a 160% improvement over baseline. The company said generation artifacts degraded results at higher synthetic-data volumes.