I'm using the CascadeClassifier object to detect faces in a video sequence, but I'm seeing a high rate of false positives. To combat this, I thought I would try to detect eye pairs within the ROI of the detected faces, but several of my videos seem to be too low resolution to do the eye detection.
What is the recommended alternative for avoiding false positives?
Are you using the default OpenCV cascades? If not, you can train you own cascade, here is the guide. And take as much positives as you can, FDDB is a good database, but you can augment it with others. You should also use a large number of negatives during the training.
Another possible way (and preferable) is to apply tracking of faces. You can keep a list of faces and compare them between frames. Than, if face is detected regularly (i.e. 5 detections in a row), you consider it as a real face. BTW, check the DetectionBasedTracker class, it probably does what you want.
One method for avoiding false positives might be to check for the size of the detected blob, or to check its aspect ratio. Faces must have a particular aspect ratio and thus eliminates false positives
Asked: 2012-06-29 09:28:19 -0600
Seen: 7,804 times
Last updated: Mar 05 '16
