Synthetic data boosts person detection mAP50 by 160%: AWS details industrial safety AI pipeline
AWS ML Blog · rss · 2026-09-17
AWS ML Blog published an industrial safety AI solution for the scarcity of training images showing people near heavy machinery: a Qwen-Image-Edit-2509 diffusion model deployed on SageMaker AI (4×A10G, 60GB weights) inserts synthetic people into real scene images, while Amazon Rekognition DetectLabels auto-generates YOLO bounding-box pseudo-labels — no manual annotation.
- Results: up to 160% improvement in person-detection mAP50, with no hazardous staged photography or manual labeling ($3–5/image otherwise).
- Why it matters: the highest-risk scenarios are the rarest in real data and unsafe to stage; edge devices need lightweight models where every sample counts.
- Key approach: editing real images rather than fully synthetic scenes avoids domain gap; structured prompts specify hazardous placement, scale and lighting constraints; Rekognition runs at 80% confidence with NMS dedup, converted to YOLO and merged with existing equipment annotations.
- Cost: 166s/image currently; projected 10x inference cost drop on H100 or with quantization (not yet validated).
More from Research
- How to Loop Your Transformer: Meta/MIT arXiv paper on loop scaling laws — rohanpaul_ai · 2026-11-02
- Meta & MIT paper: looping transformer layers beats adding depth at fixed size — rohanpaul_ai · 2026-11-02
- AI tell-tale word 'delve' declining in latest NeurIPS papers, analysis shows — JungWooHa2 · 2026-11-02
- ByteDance Seed's Loop-μP paper draws 'Chinese loopmaxxers' buzz — scaling01 · 2026-11-02
- New benchmark shows frontier VLMs still can't read Braille despite 43M blind users — prajdabre · 2026-11-02
- Diffusion models emerge from associative memory: spurious states may signal generalization — DimaKrotov · 2026-11-02