[Libface-general] A possible TODO addition

[Aditya B. <adi...@gm...>]

Hi Alex,

If you've been following the libface-related thread on digikam-devel, you
already know this, but I'd like to mention this anyway.
A professor in my university recently showed me a paper which proposes a
tweak to the eigenfaces algorithm. The idea is to use Global Linear
Regression (GLR) training to "imagine" a virtual frontal face from a single
input profile/side face. The authors claim that recognition accuracy becomes
pretty high after using this.

Do a google for H-Eigenfaces or Hybrid Eigenfaces. You might also like to
search for a paper entitled "Local Linear Regression (LLR) for Pose
Invariant Face Recognition". If explains nicely how to use GLR and LLR for
constructing a virtual frontal face. If you don't have a subscription to one
of the websites for downloading, I'll send you a copy of the papers.

I think this should go into our TODO list. If we go for GLR, we don't have
to change the eigenfaces code (which I'll commit after about a week and a
half, my final exams are next week). The GLR can be implemented as a
separate module. The detector module can feed the detected faces ( without
knowing the orientation ) to the GLR rotator module, which will then feed
the virtual frontal face to the eigen/fisherfaces module. That way, our
eigenfaces database shall only consist of frontal faces, i.e. no separate
databases for profile faces, etc.

Coupled with fisherfaces, this should achieve a good degree of both pose and
illumination invariance.
And the main advantage is that we only need to implement a single module
between the detector and recognizer, while keeping the face database
extremely simple, i.e. only frontal faces.

What do you think?
-- 
Aditya Bhatt
Blog : http://adityabhatt.wordpress.com
Face Recognition Library : http://libface.sourceforge.net