Exam AWS Certified Machine Learning - Specialty topic 1 question 137 discussion

Answer is D: A is not correct since restaurant has limited bandwidth B is not correct since cannot enable Rekognition service on DeepLens C is not correct the same reason as A

AWS will not recommend to use Deeplense in production. From https://aws.amazon.com/deeplens/device-terms-of-use/

Sorry guys, not B, C or D. Reasons? Deeplens is a deprecated product, not suitable for being used in real production environment (as clearly stated in its T&C), thus B & D option are out. Between A & C, the clearly option is A. C implies the creation of a custom ML model. Making a custom model is very expensive, time consuming, error prone and a highly specialized task. Option A uses a well-known, key-in-hand service as AWS Rekognition which implies very little effort in comparison with uses a custom-made one. I know, this option does not follow the flock, but I think that I am right.

Yes, Amazon Rekognition can be integrated with AWS DeepLens. You can use AWS DeepLens to capture video and perform initial processing on the device. For more advanced image and video analysis, you can send frames from DeepLens to Amazon Rekognition

The best solution for building a line-counting application for use in a quick-service restaurant is to usethe following steps:Build a custom model in Amazon SageMaker to recognize the number of people in an image. AmazonSageMaker is a fully managed service that provides tools and workflows for building, training, anddeploying machine learning models. A custom model can be tailored to the specific use case of linecounting and achieve higher accuracy than a generic model1 Deploy AWS DeepLens cameras in therestaurant to capture video

B. AWS DeepLens with Local Amazon Rekognition and AWS Lambda: AWS DeepLens is designed for local processing and can run models at the edge (i.e., on the device itself). This setup would enable local analysis of the video feed without the need to stream the video to the cloud, thus conserving bandwidth. Amazon Rekognition and Lambda can then be used to analyze the footage and send notifications. This option aligns well with the bandwidth limitations. D. Custom Model on AWS DeepLens with AWS Lambda: Deploying a custom model built in SageMaker to AWS DeepLens allows for local processing of video data. This option also avoids the bandwidth issue by processing data on the device. However, developing a custom model might be more complex than using pre-built solutions like Amazon Rekognition.

Rekognition is a managed service. It uses API's and can't be deployed locally on devices. What we need here is local inference on the camera. AWS DeepLens comes pre-installed with a high performance, efficient, optimized inference engine for deep learning using Apache MXNet.

I would go with A, As DeepLenght is not for production workloads, we are left with A or C. A requires less effort.

B : https://aws.amazon.com/ko/blogs/machine-learning/building-a-smart-garage-door-opener-with-aws-deeplens-and-amazon-rekognition/

Based on the requirements, the best solution is option D. This option uses AWS DeepLens cameras to capture video and process it locally on the device, without sending any video streams to external services. This reduces the bandwidth consumption and avoids impacting other operations in the restaurant. The option also uses a custom model built in Amazon SageMaker to recognize the number of people in an image, which can be more accurate and tailored to the specific use case than a generic face detection model. The option also deploys an AWS Lambda function to the cameras to use the model to count people and send an Amazon Simple Notification Service (Amazon SNS) notification if the line is too long.

it is D. "The restaurant locations have limited bandwidth for connections to external services and cannot accommodate multiple video streams without impacting other operations." So, using Amazon Kinesis Video Streams is not a solution here. Ok, DeepLens dissapears in 2024...but this questions is for 2022... In the real world, the restaurant would buy good signal internet and use answer C, which is better solution.

C is Answer

AWS DeepLens will reach end-of-ilfe in 31/01/2024 so, I don't think this question will even appear in the exam.

Deeplens + lambda + model inference

After giving this some thought, I am thinking D. Tricky, my initial answer was C. But D is a better solution - given DeepLens and counting the number of people.

https://aws.amazon.com/ko/blogs/machine-learning/building-a-smart-garage-door-opener-with-aws-deeplens-and-amazon-rekognition/

https://aws.amazon.com/blogs/machine-learning/optimize-workforce-in-your-store-using-amazon-rekognition/ B