I don't mean to play it down, but honestly classifying objects and identifying crowds don't seem like super valuable outcomes from such a huge investment. I guess I expected something like "persistently label and narrow down the definition of an object over the span of multiple occasions" or "identify the intent or demeanor of a crowd".

They are using this to routinely pull wanted criminals out of crowds using facial recognition. Seems valuable especially to somewhere like China that likes to grab particular people.

Do you realize if you did this at scale how much faster you could look through recordings ? And that this is a manual process currently? Also, they are just starting out .