Comment 77 for bug 507511

http://freshmeat.net/projects/kolmogorov - Seems to have been moved somewhere.

I had an idea. What if we use the classifier that comes with OpenCV as a base, get the rectangle where the face is. Then use this much smaller image to identify the face. We can use those small images to train classifier in OpenCV. In effect current face detect classifier in OpenCV is as general as it can be, to recognise a generic face, so what we will be creating is a highly specific classifier tailored for each face in the training data, each face in our gallery.

Now, what is needed is to improve general OpenCV classifier locally, because for example, it won't recognise some faces which are altered in spacial domain, e.g. turned 90 deg left/right. I tested this on my machine with an image of Lena (popular test image), which it happely recognises in its original form, but rejects if rotated 90 deg. I had a look at the source code for the haartraining and it needs to be modified slightly because it creates a fresh classifier from scratch everytime, where as what we want is one that takes an existing classifier, then goes through the same motion to update it internally.

Advantages of using OpenCV, as a start, is because it already has lots of features in its bank. However, if that doesn't work well then we can think of an alternative way. I hope to get my hands on some image processing books and papers once I am back at university from holidays. :)