Peter M. Groen / openbr

Eigenutil cleanup

Eval landmark improvements

Block based model compression scheme

…r into eval_landmark_improvements

Conflicts:
	openbr/core/eval.cpp

Switch FaceRecognition to UnitDistance instead of
MatchProbability--there are some doubts about using gaussians to model
the score distributions, and the KDE code is untested.

Remove FTETransform from age/gender classification (again, it is
unused). Drop a couple extraneous Expands.

Parallelize KDE lookup table building, also introduce a flag, and
don't do KDE training at all, if Gaussians will instead be used to
model the score distributions

Also address an issue with retraining algorithms, using pre-trained
submodules

Error message for ReadTransform

Blur regions of interest

Implemented RndSample transform