It's extra work, but is more contextual. You could, for instance, link to wikipedia articles, imdb articles (for actors) etc. and not just link to ad content.

Really what it does is help define what is inside an image.

Although I despise ads, I have to admit your tech is nice.

Do you do automatic image segmentation, or are these images "hand-labeled" (I can not imagine that would scale)? If it's automatic - is it calculated server-side or client-side?

They're hand-labeled. In the future, I'm sure I'll be hiring a bunch of Google people to build the "automatic" part.

This should be something that will improve over time with machine learning and can already be done right now.

Not a lot of extra work when compared to the amount of work that goes into existing advertising campaigns.