[mlthoughts2018]

Surveillance video processing and intelligence services are probably by far the largest use cases for video background removal and similar techniques. Far smaller in terms of number of customers, but far far larger in terms of revenue specifically for this capability.

Millions of people paying peanuts (attributable to this feature) is way less than a few hundred or thousand research / defense labs, contractors, etc., paying tens of thousands up to even millions on licensed software for this type of thing.

[rasz]

In both teleconferencing and surveillance you are dealing with static background, no need for fancy ML algorithms.

[mlthoughts2018]

No, not true. Surveillance cameras mounted on police cars, helicopters, etc. Even in stationary cameras, scene understanding is a hard problem. Suppose someone wheels a dolly of boxes into a warehouse and leaves them. Are they now part of the background? Are they suspicious boxes? The coarse level state transition of objects in the scene is really hard to solve even with heavy ML.