- Saliency stl array instead of pointer array added
- Perf test in Extra Repository
- Epsilon change to 10E-3

predict method rewritten to use with PredictCollector interface (abstract class)
now face recognizer allow to perform different scenarios of result handling

Issue 1:
The subpixel precision of LSD was lost in the implicit conversion to Vec4i
Solved:
Line point coordinates are stored in cv::Vec4f
No type conversion needed from extremes to kl float fields

Issue 2:
The keyline point fields for original image were not being correctly filled
Solved:
The points in original image are computed through scaling (using scale factor and octave index)

results in less code and more consistent parameter handling.