The process of marking faces as specific people is partially manual, and so has the potential for user error, with a face assigned to the wrong person. This would make the face recognition model less accurate.
Would it be possible to use the model to look at already assigned faces, and find those that are furthest from them model? (Since these are likely to be wrong)
For example, use the model for the person ‘Mike’ to determine if any faces marked as ‘Mike’ do not have much probability of actually being ‘Mike’.
I can see a few challenges though:
- I could have a picture of the back of someone’s head, and I have marked it as ‘Mike’, even though the face isn’t really visible. I would suggest the best way to deal with this is to only apply this to faces that the face detection is able to detect automatically.
There might also be cases in which I have marked a blurry picture as ‘Mike’, when I know who the person is, but the face isn’t programmatically recognizable.
