61-- / klucv2 Goto Github PK

What steps will reproduce the problem?

1. Choose and image with a face in which the mouth is opened.
2. Watch what is detected.

The mouth is expected to be detected correctly.
Instead you see either the space between the lips detected or some other shape.

Maybe we should ensure that the left and right corner of the mouth are only
accepted if they lie left and right from the anthropometric mouth point.

For ANN-data normalization it might be relevant how a face is rotated
around the z-axis (your perpendicular viewing axis on the face).

Maybe we should rotate the face "back" using the negative angle of the
connection line between the two eyes.

Currently the eye region is recognized by a cascade of Haar-based
classifiers...

When detecting the eye region, the eye ofton does not lie completely in the
region's rectangle. Sometimes a corner is cut away.

Maby try some other haarcascade (for instance the ones from Modesto
Castrillon-Santanaones) to fix it:

http://code.google.com/p/klucv2/source/browse/trunk/data/haarcascades/haarcascad
e_mcs_lefteye.xml
and
http://code.google.com/p/klucv2/source/browse/trunk/data/haarcascades/haarcascad
e_mcs_righteye.xml

and

http://code.google.com/p/klucv2/source/browse/trunk/data/haarcascades/haarcascad
e_mcs_lefteye.xml
and
http://code.google.com/p/klucv2/source/browse/trunk/data/haarcascades/haarcascad
e_mcs_lefteye.xml

I expect the detected eye rectangle to exactly fit the eye at least when
looking at the camera in frontal view.