Hey guys, I'm having an issue with an ML problem. Essentially I'm trying to code a facial "verification" system. Typically face recognition models use 1:N matching which means;
- We have a set of N identities
- Given image, which identity is represented in the image?
I'm trying to do a 1:1 verification system, which works like this:
- We have one identity
- Given image, does this image depict the identity?
You can see that the first system assumes a match occurs and tries to pin it down to the correct person, whereas the second system is checking if a match exists.
The difficulty is this -- ML does not work very well for this problem. Typically a classification algorithm uses training data which attempts to describe the difference between classes (e.g. this information represents THIS class, that information represents THAT class, etc), and this is analogous to the 1:N face recognition task. Face verification, on the other hand, only uses a training set with one individual class (there's no class split to discern) and attempts to "learn" to identify that class, but since it doesn't have contrasting information it's not very good at doing that.
My question is this: what kinds of techniques, if any, exist for identification instead of classification? If I can't find any good techniques my tentative idea is to mask the 1:1 verification as a 1:N matching problem (so essentially I'll just get a bunch of random face images and collectively describe them as the "other class"), but this requires a very large data set that I don't currently have, so I'd like to avoid that option if possible.
[link][5 comments]