AWS Rekognition and Sagemaker Running Local to automate annotation base on video streaming

Amazon Monitron could be a good addition to consider in this scenario. Monitron is an IoT platform that includes a suite of sensors, gateways, and machine learning algorithms that help industrial customers monitor the health and performance of their equipment. Monitron can be used to automatically detect equipment anomalies and predict when maintenance is required. There's actually also a whole suite of other solutions and services for industrial.

In your case, Monitron could be a complementary solution to implement along with Amazon Rekognition and SageMaker. You could use Monitron to collect sensor data and use Rekognition to process the video data and SageMaker to train the machine learning models. Or you could you use Monitron by itself, as it is an end-to-end system.

A knowledge of AWS services such as Amazon Rekognition, SageMaker, and Greengrass would help. For studying, https://workshops.aws/ has some good hands-on training on some of these topics. It would also be helpful to study Monitron in more detail and understand how to integrate it with the other AWS services you plan to use.

Here's a blog that talks about using Monitron to reduce unplanned downtime: https://aws.amazon.com/blogs/aws/amazon-monitron-a-simple-cost-effective-service-enabling-predictive-maintenance/

I would also recommend you also get in touch with your AWS account team. They may be able to provide some time with a Solutions Architect to dive a little deeper into helping you think through the architecture.

If you can have connectivity then Amazon Rekognition Photo Custom Labels can be a good start, by sampling a live video stream for annotation event - this would be the simplest solution but with less flexibility in terms of personalization of your model.

Otherwise, if you want what your model to work offline or you want to train a custom Computer Vision model, the Greengrass for the inference but leverage the power of the SageMaker + GPU ML instances for training your model in cloud. For the later approach, please try the Quick Starts in SageMaker Studio that you can try for Image Classification sample notebook: https://github.com/aws/amazon-sagemaker-examples/blob/main/introduction_to_amazon_algorithms/imageclassification_mscoco_multi_label/Image-classification-multilabel-lst.ipynb Note that in this case you can leverage the SageMaker framework to train faster, optimize the hyperparameters using build-in tools like "hyperparameters tuning" feature: https://docs.aws.amazon.com/sagemaker/latest/dg/automatic-model-tuning-how-it-works.html

A 3rd option could be Amazon Lookout for Vision agent for Greengrass if you want only to detect a defect: https://docs.aws.amazon.com/greengrass/v2/developerguide/lookout-for-vision-edge-agent-component.html