Linux Photo Tools (The H)

Posted Jun 8, 2011 15:05 UTC (Wed) by knobunc (guest, #4678)
Parent article: Linux Photo Tools (The H)

I'm still looking for a way to do face detection and recognition under Linux.

OpenCV looks promising, but I've never managed to scrounge the time and inclination to try it since it provides hooks, but no UI for it.

Linux Photo Tools (The H)

Posted Jun 8, 2011 15:12 UTC (Wed) by milliams (guest, #71641) [Link]

I believe that the new DigiKam (2.0) will provides it when it is released (http://adityabhatt.wordpress.com/2010/05/19/yet-another-g...).

Linux Photo Tools (The H)

Posted Jun 8, 2011 23:52 UTC (Wed) by endecotp (guest, #36428) [Link] (4 responses)

Using OpenCV for face tracking is apparently not difficult:

http://www.statusq.org/archives/2011/05/15/3407/

Linux Photo Tools (The H)

Posted Jun 8, 2011 23:59 UTC (Wed) by knobunc (guest, #4678) [Link] (3 responses)

Yeah, tracking is good. But I want recognition (so that I don't have to say who is in the pictures).

Linux Photo Tools (The H)

Posted Jun 9, 2011 15:35 UTC (Thu) by dashesy (guest, #74652) [Link] (2 responses)

Face recognition is a much more difficult problem than tracking, and is generally known as cognition.

First, the image should be registered (you may avoid this if you have a very good quality image with enough resolution and the person has not moved her face in an awkward angle)
Second, features are extracted (to reduce the dimension); you may want to get some heuristic approach for this, or rely on a statistical approaches or on some eigenvalue decomposition, or wavelet transformed (thus reduced) feature space, or a combination.
Third, you have to perform steps 1 and 2 for a set of people you are interested in, and record the features in a database
Fourth, when you have a test image (to recognize the face) you can again perform the first two steps then compare the result against your database.
If your feature extraction is hierarchical with low-to-high (fast-to-slow, coarse-to-fine) approach then recognition may be fast enough to be practical.

Now, with OpenCV you can do all of the above!
And if you have an Intel CPU you can get the best from OpenCV using IPP.

Linux Photo Tools (The H)

Posted Jun 9, 2011 16:55 UTC (Thu) by knobunc (guest, #4678) [Link] (1 responses)

If you have any good pointers for books or articles or documentation explaining OpenCV to a relative lay person that would be great.

I followed the DigiKam lead and found libface that may do some of what I want. I had to fix a couple of bugs in the sample code, so I haven't managed to see how good the software is at recognizing people yet.

Linux Photo Tools (The H)

Posted Jun 9, 2011 18:25 UTC (Thu) by dashesy (guest, #74652) [Link]

It might be a good idea to read the first few chapters from the classical Learning OpenCV, and that is more than enough because more likely no body needs all the features. Since the notation and class hierarchy is more important I found the cheat sheet more useful :) I have done pattern matching (not for face actually) and cannot tell anything regarding the performance of libface. Since there are many ways to extract the features you should find one that meets your needs regarding the speed and quality.