Many computer vision applications use video source data, but treat the task as independent analysis of a sequence of images - for latency, resource cost, or project complexity reasons.
There are lots of ways to convert a video to a set of images, but the Python binding of OpenCV is one popular tool for its flexibility: Different projects might need to e.g. extract at different frame rates, crop the results, or combine with other transformations.
AWS also provides a range of differently-optimized environments on which the code might be run, such as:
- As a Lambda function (like this greyscaling example), which could be triggered to run automatically whenever a new video was uploaded to S3)
- As a SageMaker Processing Job, well-suited for on-demand batch processing for data scientists who want scalable compute without leaving the SageMaker workbench.
This example isn't intended as a reference implementation of OpenCV frame extraction logic - but it demonstrates some patterns for using SageMaker Processing that might not be obvious to beginners, like:
- How to use dependency libraries that aren't available in the SageMaker base container images (and especially those like OpenCV that might require extra OS-level libraries beyond a simple
pip install) - How to pass parameters in to a processing job (like frame rate, which we pass in but don't implement in the container logic)
- How a processing instance can discover the total number of instances executing the current job, and its place in the list, since this process is a little different than the
SM_HOSTSandSM_CURRENT_HOSTSenvironment variables often used in SageMaker training jobs.
Note that this use case doesn't need to identify instance count (because SageMaker can distribute input data between instances automatically with the s3_data_distribution_type parameter), but the logic may be useful for other applications e.g. master-slave cluster architectures.