How to Build This | S1E3 Getting Started with Computer Vision in AWS

Start building on AWS today!

Watch how you can build a computer vision app at scale while achieving low latency. You’ll learn how you can train a model in the cloud and deploy it to edge devices. You’ll also explore how to evaluate what hardware to use for your computer vision app. I’ll walk through an example of using computer vision to improve the accuracy of line cooks putting food deliveries in the right bag.


Special thanks to these contributors:
* Britton Winterose
* Mike Apted
* Dylan Qu
* Mike Tuszynski
* Austin Ashe
* Mo Mobarak
* Brintha Koether
* Emily Webber

Resources for this episode:
* Find AWS IOT Core and AWS Greengrass compatible cameras in the partner device catalog:
* How to integrate NVIDIA with AWS IOT Core and AWS Greengrass:
* Using your own algorithms and apps in Amazon SageMaker:
* How to use Amazon SageMaker Groundtruth:
* Evaluating your model in Amazon SageMaker:
* Deploying models to Amazon Greengrass:
* DynamoDB Streams use cases and patterns:
* Using AWS Lambda with AWS IOT Greengrass:
* Getting started with Amazon Augmented AI:

AWS partners that produce thermal cameras:

Consider these points when picking a camera:
* Budget? Cameras can range from $10 to $1M each depending on application/use case
* Where will the camera be? I.e. indoor/outdoor
* What environment will it be exposed to? High temp industrial setting? Commercial setting? Residential? How will the camera be mounted?
* Does it need to be serviced/how often?
* Desired life expectancy/warranty
* Is the goal for the camera to be dual purpose (image viewing, video, and computer vision)? Or Computer vision only?

Use case specific (i.e. personal security, situational awareness, manufacturing quality control, event based recording, inspections, etc.)
* How much light is available on the objects being detected?
* How many cameras are needed?
* Do they need to operate independently or together (AKA “stitched”)
* How much latency is required for the output of the CV data? Is local storage needed?
* Connectivity: how will the camera transmit its data to the cloud?
* How close will the objects be from the camera being scanned? This will help determine things like resolution, framerate, and bitrate
* Does the camera need to Pan, Tilt, or Zoom? If so, remotely?
* Is color important?
* What types of detections are needed? This will determine CPU/VPU/GPU and any hardware acceleration requirements and what type of CNN are needed (if any)
* What level of accuracies are needed?

* Who will do machine learning modeling/training?
* Where will the model live?
* Who will annotate the training material?
* How many detectors are needed to run simultaneously?
* How much back end support will be needed and who will support it?

Duration: 00:10:01
Publisher: Amazon Web Services
You can watch this video also at the source.