Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 176

Fig. 18 Baseline for profile face model
O PEN CV provides a Haar cascade only for left-facing profiles. Normalising all
training samples to the same view is a more efficient use of the data. In order to
detect right-facing profiles, the images must first be horizontally flipped.
As previously discussed, using the O PEN CV face detector in profile view is not
as reliable as frontal view, because the classifier relies on block features. It is inter-
esting to note that the ear in frontal view is an outline feature, but in profile view
it becomes a block feature. It would therefore make sense to use the ear as a key
feature in profile view. Ear detection using cascaded Adaboost from a profile view
is discussed in [12].
If the O PEN CV face detector is trained with profile faces, the classifier has to
attempt to learn the background variability beside the profile edge. The O PEN CV
Haar cascade provided for profile faces works reasonably well against a plain back-
ground, but to provide reliable detection against variable backgrounds, it would need
to be trained with a much larger data set with highly random backgrounds [5].
Viola-Jones requires a very large set of training data. There would need to be
thousands of positive examples of faces, and tens of thousands of non-faces. When
training a cascade, there should be many more negative examples than positive