Here's a thought - why don't face recognition systems not only find a match in their database but like many recommender systems, find nearby matches (netflix recommends me movies and says "because you watched this and liked it") but also, without being privacy invasive, one could just use synthetic faces and the error (precision/recall) in the face recognition network, to show the range of kinds of faces that would also match - e.g. can just use a GAN like this fake face generator site does:
Note propagating error information is also already the basis for all those fancy stable diffusion image things like Dall-E
A nice explainer here, for example: so should be very easy.
A neat thing to do with this would be to virtualise "identity parades". - instead of picking some likely looking suspects off the street, one could generate a set of faces (or figures or even videos)....one thing this would also do would be to underline how bad humans are at identifying people from a remembered incident - indeed, one could do some interesting cognition/perception/memory/bias research on the differences between poorly trained AIs, and poor old people.
We could call this "facebookem" perhaps :-)
A nice demo would be a "dark mirror" app, that would run on your camera phone, and pick at random a synthetic face that would pass for you in a line-up.
This paper on face replacement does a similar job, but uses different faces, rather than ones drawn from a GAN trained on the face we're trying to undetect.
No comments:
Post a Comment