How were the stock face detectors trained/tested?

asked 2013-12-04 13:35:53 -0600

xaffeine gravatar image

updated 2013-12-11 18:36:55 -0600

The OpenCV distribution contains several xml files describing classifier cascades for frontal face detection. Can anyone tell us what database(s) and methodology were used to train and test them? Specifically, I am thinking about haarcascade_frontalface_alt.xml, haarcascade_frontalface_alt2.xml, haarcascade_frontalface_default.xml, and lbpcascade_frontalface.xml.

Is it a secret?

edit retag flag offensive close merge delete

Comments

Somebody must know the answer. Don't they need to get retrained whenever related code changes?

xaffeine gravatar imagexaffeine ( 2013-12-17 13:30:50 -0600 )edit

I don't know the original process but the links given below might give you some idea.

http://docs.opencv.org/doc/user_guide/ug_traincascade.html

http://coding-robin.de/2013/07/22/train-your-own-opencv-haar-classifier.html

prakharmohan gravatar imageprakharmohan ( 2014-02-20 05:58:46 -0600 )edit

I am not asking how I could do training; I am asking for details on how the particular models in the OpenCV distribution were trained. It seems strange to me that an open-source project would keep these details secret.

xaffeine gravatar imagexaffeine ( 2014-02-20 12:39:03 -0600 )edit
add a comment