So i'm in the process of increasing face recognition by adding preprocessing (CLAHE, Tan-triggs, Equalizehist) and hoping that will give me better results. But looking at some posts here and there, the results are not as good as the whitepapers make them out to be. I am using the fisherface method.
I get good enough results currently, but not good enough for a stable recognition (people get recognized as other people for example) and the difference between known and the unknown threshold is usually to little use that. Currently my eigenvalues are are between 150-200 with good lighting (or good training pictures taken) but go to 400 when the situation is worse. It doesnt matter if i set the threshold higher, as sometimes i still get a picture that is unknown to be recognized as a known (as mentioned previously).
I will also try alignment (congealing code) as linked to here: https://bitbucket.org/gbhuang/congealreal Has anyone succesfully used this with opencv3 ? Does it work ?
My question is; what is really the route to getting superlow eigenvalues so i could use it in a production situation (i.e. authentication for example). I don't see any real tutorials on how to get this perfected, i realize this is a hot topic and maybe no one wants to spill the beans on what worked for their product :-)
Tantriggs for eliminating lighting issues, alignment and flandmark ?
Appreciate anyone's comment to even point me in the right direction. Thanks, atv