Haar cascades for eyes?

[Horea Christian]

Hi guys, I'm currently trying to write a script for image post processing (I want to create a well-documented and standardized open face database). The haar cascades which come with openCV are awful (or I just don't know how to use them). They give me lots of false positives/negatives. Also they are quite bad at telling left from right eyes.

Do you know where I can find better haar cascades? I'd either like them to be separate for left/right eyes or to be able to output two eye fields (I want to rotate images so that the eyes are on one horizontal line, so I need two separate references). If you want to have a look at how I'm implementing the haar cascades, my code's on github:
https://github.com/TheChymera/Epproc/commit/b1e649be19c425e6f284beff0b637211f73892a8#diff-4

It's pretty much my first time with computer vision so any help at all would be appreciated!

Cheers,

[David Bolme]

I have found that the haar cascades for frontal face detection works pretty well when compared  to other face detectors although commercial face detectors are often more robust under difficult lighting conditions.  The eye detectors are unpredictable though.  It is possible to determine the eye coordinates relative to the face bounding box to get an initial guess.  You can then select the cascade detection nearest to the expected locations of the eyes.  I also have a correlation based eye detector in pyvision that is based on my dissertation work.  There is a simple tutorial you can access on the pyvision website that will show you how to set it up.  I believe it even shows how to use an affine transform to align the eyes.  See the link below.

http://sourceforge.net/apps/mediawiki/pyvision/index.php?title=Face_Detection_and_Eye_Localization

---
Dave

--
 
---
You received this message because you are subscribed to the Google Groups "pyvision" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pyvision+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.