Hello,
I am trying the facerec_demo.py script on the AT&T face database and the confidence value returned for any image in the original training database is 0.00. Is that expected behavior?
Thanks!
patrick
well, oh, the python demo is a bit flawed there.
the image you want to test should be excluded from the training set ( like the cpp demo does )
since this does not happen, think of it, it's no wonder that it finds zero distance to itself
Exactly what he says. You should always divide your data into three sets: training, validating and test set. In order to know what the meaning of both is, take a look at this link: http://stats.stackexchange.com/questions/19048/what-is-the-difference-between-test-set-and-validation-set. It gives a good explanation of the basic concepts behind training correctly and how to reach the highest accuracy, while avoiding overfitting.
Thanks @berak for the reply. And yes, I am familiar with cross-validation methods. So it seems I don't understand the meaning of "confidence" as used with the FaceRecognizer. Are you saying that the meaning of "confidence" is the same as "distance"? This seems backwards to me. If the distance between the test image and itself is zero, shouldn't the confidence of the match be 100% (1.00)?
Asked: 2013-04-22 11:51:43 -0600
Seen: 233 times
Last updated: Apr 22 '13
