@elight I think that generally requires a large enough data set to do something with.
This is the problem I ran into building face recognition for my photo library: I just don't have that many pictures, and especially not enough carefully aligned portraits, to build real training sets.
So anything I do alone is constrained by outside sources.