AWS has presented a reference architecture for managing computer vision models across large camera fleets with the current Amazon SageMaker Python SDK v3. The company says the approach is designed for deployments where differences in camera angle, glare, and lighting make individually fine-tuned models more effective than a single model trained on pooled footage.

The architecture uses a parameterized Amazon SageMaker Pipeline for each camera. Its two training branches produce an anomaly detector based on a convolutional autoencoder and a ResNet18 defect classifier. Each branch exports its model to ONNX, evaluates the exported artifact against held-out data, and applies a quality gate. Models that pass are compiled with SageMaker Neo for the configured edge target and registered in the SageMaker Model Registry. AWS says the resulting packages include portable ONNX models and Neo-compiled, device-optimized artifacts, with camera identifiers used to resume training from each camera’s latest approved model.

An orchestration layer uses Amazon SQS, AWS Lambda, AWS Step Functions, and a DynamoDB semaphore to control concurrent retraining. The sample admits up to five pipeline executions at a time by default. Completion and failure notifications are sent through Amazon SNS, with operational failures routed by Amazon EventBridge.

AWS says the sample is deployed with the AWS Cloud Development Kit and that a full pipeline run on synthetic data costs about $0.07 in compute. The architecture replaces the deprecated Amazon SageMaker Edge Manager with SageMaker Neo. The sample is available in an AWS samples GitHub repository.