AVA: The Art and Science of Image Discovery at Netflix

Actually, this might be one of the intended effects of this feature. What better way to make it appear their catalog is wider and ever-changing then to cycle through these static stills every so often. It may not be the most user friendly design pattern to 'trick' your user into thinking there's a bigger catalog, but I can understand the desire for it from their perspective.

And to be perfectly honest, sometimes those changing stills have actually made me watch a movie that I had previously glossed over because the image wasn't as interesting. So from personal experience I can say their primary intent was reached w/ the periodic cycling of stills.

This is an incredible product.

I have come across the github repo's authors work a number of times now, and I am continually impressed by the quality of documentation, examples, and immediately usable code.

They probably do, for efficiency. You have to process a frame to some extent to detect motion blur or darkness (how else could you possibly know?), but that doesn't mean you can't abort the processing as soon as you detect the frame is totally unsuitable. I expect whole groups of frames may be able to be discarded more-or-less together, e.g. a dark key frame followed by frames with minimal difference, or where the video encoding's motion estimation says there's a lot moving.

Netflix cancelled its second million-dollar Netflix Prize in 2010 due to privacy lawsuits. Netflix has not outsourced its data science since.

https://www.wired.com/2010/03/netflix-cancels-contest/