Amazon SageMaker AI pipeline lifts person detection 160%

Amazon's post walks through a synthetic data augmentation pipeline that generates photo-realistic training images by editing real photos of heavy machinery with a diffusion model, then auto-labeling the inserted people with Amazon Rekognition. The approach targets a familiar bottleneck in industrial safety AI: the highest-risk scenarios, such as a person standing in a vehicle's blind spot, are the rarest in real datasets and the most dangerous to stage. By inserting synthetic people into existing equipment photos rather than generating entire scenes from scratch, Amazon says the method preserves background fidelity and keeps existing equipment annotations valid.

In experiments, Amazon used an OpenImages subset of about 3,200 real train images and up to 1,000 synthetic images, then trained YOLO11 detectors. The results showed that where a person is placed in the prompt mattered more than how the scene was varied: hazardous placement doubled person detection mAP50, while background placement slightly degraded it. Performance peaked at 750 synthetic images before declining, and the medium-capacity YOLO11 model achieved the best aggregate result. The report positions the pipeline as a cost and safety win, with per-image synthetic cost estimated at $0.33 versus $3–$5 for manual collection.

The stakes hinge on whether these synthetic gains transfer to other domains and production edge deployments, since Amazon's experiments used trains as a publicly shareable proxy for heavy machinery. The company also notes it has not yet validated its projection that inference cost per image will drop around 10x on more capable hardware. For safety teams, the takeaway is that prompt-guided placement strategy and synthetic volume may become tuning choices as consequential as model architecture.