[candiodari]

I don't get this. This is about tracking humans. If you're worried about AIs tracking humans through areal cameras ... let me put your mind at ease : this is "tutorial" level stuff.

Granted. There's getting it to demoable state, and there's getting it to work under all conditions, getting it stable, getting it tested, and so on and so forth.

But still, this is not exactly state of the art anymore. This ship has sailed. Over and done. Genie cannot be put into the bottle. The US army has this option now, and very soon essentially any professional military will have it. A quick course on AI will enable you to do this, and I assume that the US military has enough such people available.

Same with tracking specific people in (high-res) cams. There's a computational cost, but this has been done and described so many times. If anybody wants to build a network of cameras that can track specific people by their faces, there's nothing stopping them at this point.

So why get all worked up about this ? What's the big deal ?

[jononor]

If the US army had or was easily able to build this they would probably not pay large sums of money to Google. There are still many challenges in creating scalable robust solutions using machine learning in computer vision, especially when high degrees of response time and robustness is needed.

[candiodari]

Most, if not all types of machine learning networks are O(1). Keeping the response time constant is not needed.

Robustness is harder, but for this problem, not very hard.