Enhancing industrial safety AI with synthetic data on Amazon SageMaker AI

來源原標題

Enhancing industrial safety AI with synthetic data on Amazon SageMaker AI

作者

Dimitri Voytan

原始標籤

Advanced (300) · Amazon SageMaker AI · Technical How-to

原始連結

https://aws.amazon.com/blogs/machine-learning/enhancing-industrial-safety-ai-with-synthetic-data-on-amazon-sagemaker-ai/

摘要原文

Industrial safety AI refers to the use of technologies like computer vision and predictive analytics to find and stop workplace dangers. Synthetic data augmentation is emerging as a practical solution to one of the hardest problems in industrial safety AI: the scarcity of training images depicting people in dangerous proximity to heavy machinery. Industries deploying autonomous equipment (for example, agriculture, construction, mining, and manufacturing) need reliable person-detection models. However, the highest-risk scenarios (for example, a worker standing in a blind spot, a child near a moving implement) are precisely the ones that are rarest in real-world datasets. They are also the most dangerous to stage for data collection. In this post, we show how to use a synthetic data augmentation pipeline built on Amazon SageMaker AI and Amazon Rekognition to generate photo-realistic training images with automated labels. Our experiments showed up to 160 percent improvement in person detection mAP50 (mean Average Precision at an Intersection over Union threshold of 0.5) without manual annotation or hazardous photography sessions. Industrial companies developing AI-powered safety systems for heavy machinery face a critical bottleneck. They need thousands to millions of training images showing people in dangerous positions near equipment to train object detection models that can help prevent accidents. However, this data is extremely difficult to obtain: These challenges leave edge-deployed models with insufficient training signal for person detection in the exact scenarios where detection failures have the most severe consequences. This section describes our end-to-end synthetic data augmentation pipeline, which operates in two stages: photo-realistic image generation and automated labeling. Our solution is an end-to-end synthetic data augmentation pipeline that generates labeled training images without manual annotation. The pipeline operates in two stages: In our approach, we edit real images rather than generating entirely synthetic scenes from scratch. This preserves background fidelity, avoids domain gap, and allows existing equipment annotations to remain valid. Domain gap is the performance drop that occurs when a model trained on one data distribution, such as fully synthetic scenes, is applied to real-world images. The diffusion model’s scene understanding produces contextually coherent insertions with realistic lighting and proportions. We deploy the Qwen-Image-Edit-2509 diffusion model on Amazon SageMaker AI using an ml.g5.12xlarge instance (4× NVIDIA A10G GPUs, 96 GB total VRAM). The model receives a structured prompt specifying: The model edits real images containing equipment but no people as substrates, inserting synthetic people while preserving the original scene context, lighting, and scale. This in-place editing approach avoids the domain gap issues that plague fully synthetic scene generation. Each generated image is processed through the Amazon Rekognition DetectLabels API with a minimum confidence threshold of 80 percent. Detected bounding boxes are deduplicated by using non-maximum suppression (NMS) with IoU threshold > 0.5 and converted to YOLO format. These pseudo-labels (machine-generated annotations produced automatically in place of human labeling) are merged with existing equipment annotations from the original image. This automated labeling alleviates manual annotation for synthetic images while maintaining label quality sufficient for training edge-deployed detectors. The synthetic images and their automated annotations are combined with real training data to form the final training set. Pre- and post-processing steps (image filtering, image resizing, prompt assembly, and label processing) allow the pipeline to run as an automated workflow. This section covers the model deployment configuration, the synthetic image generation methodology, and the automated annotation pipeline. The Qwen-Image-Edit-2509 model is deployed on Amazon SageMaker AI with the following configuration: The model requires a custom device map that distributes its 60 transformer layers across the available GPUS. For efficient inference, we recommend avoiding cross device communication. Deploy on a single gpu with sufficient VRAM, such as the NVIDIA H100 GPU provided by ml.p5.4xlarge instances, or employ weight quantization techniques. We project that inference cost per image will drop around 10x using such hardware, but this measurement is not yet validated. The pipeline selects real training images that contain locomotives but no people as editing substrates. For each image, the diffusion model receives a structured prompt. Here is an example prompt structure: Our experiments revealed a critical finding: prompt-guided placement is the single most important factor for downstream detection performance. Instructing the model to place people in domain-relevant hazardous positions doubled person detection mAP50, while placing people in the background actually degraded performance. Each generated image is processed through the Amazon Rekognition DetectLabels API. The model computes bounding boxes for humans using the query words: person, human, man, woman, boy, girl, child, adult, people. This automates label creation at scale. We use: We used an OpenImages subset as a publicly shareable analog to our customer’s industrial equipment dataset: We evaluated the impact of synthetic data augmentation across three dimensions: prompt placement strategy, synthetic data volume, and model capacity. We evaluated our pipeline using YOLO11 as a representative edge-deployable detector family, trained on Amazon SageMaker AI with PyTorch 2.1.0. Models were trained with image size 640, early stopping (patience 50), and a maximum of 500 epochs. Our prompt ablation study with YOLO11-nano compared four synthetic data conditions against a real-data-only baseline. We used 1,000 synthetic images in each scenario. The following table shows key performance metrics for each scenario. Takeaway: What the model generates (a person in a domain-relevant position) matters far more than how the scene is varied around it (for example, ambient condition or time of day). Hazardous placement doubled person mAP50 (from 0.051 to 0.106), while background placement slightly degraded it. To explore the impact of synthetic data volume on detection performance, we swept the number of synthetic images from 250-1,000 while keeping the person placement condition and model size constant (hazardous placement with no scene variation, YOLO11-nano). The following table shows key performance metrics for each scenario: Takeaway: In our experiments, performance peaked at 750 synthetic images with 160 percent improvement in person mAP50 over baseline (from 0.051 to 0.134). Beyond this point, accumulated generation artifacts (for example, garbled faces and over-saturation) introduced noise that degraded both localization confidence and detection coverage. Treat synthetic data volume as a hyperparameter to validate, not maximize. To explore the impact of model capacity, we swept YOLO11 across five model scales with 1,000 synthetic images (hazardous placement with no scene variation): Takeaway: YOLO11-medium achieved the best aggregate performance (0.604 mAP50) and the highest person recall (0.340), doubling recall compared to the baseline. The 57M-parameter XLarge model regressed below the Nano+synth result, demonstrating that with only approximately 4,200 training images, there is insufficient data to regularize a model of that capacity. The combined pipeline improvements are particularly significant for safety: The synthetic data pipeline delivers significant cost, safety, and speed advantages over manual data collection. In the following sections, we compare the two approaches, outline scalability benefits, and summarize time-to-deployment improvements.