Correct me if I'm wrong, but this is just a frame-by-frame labeling. You can download whatever pre-trained CNN, pass individual frames through it and get the same result.
At the very least in this release they are identifying scene changes, which is a dynamic property.
The value prop is similar to, say, Twilio. Though, arguably, it's easier to run your own pre-trained CNN than it is to replicate the telephony, VoIP, and video conferencing stuff Twilio provides.
Also, presumably Google is hoping that they can continue to train and improve their CNN so that it's always just a little better than the best free-to-download ones.
There're papers like [1] where CNN output is used as input for RNN, which performs deeper context analysis. Results aren't exciting, though.