tion of facial points) is helpful to distinguish faces/non-faces. In Section 2,
we experimentally verified that a simple SVM classifier using facial point based
features in the post processing can significantly improve the accuracy of a ViolaJones
detector. This finding is not surprising because face alignment finds corresponding
parts between faces, makes them directly comparable, and simplifies
the face/non-face classification problem. Similar observations have also been
made in [32, 14]. While the observation is clearly useful, the real problem is how
to use it effectively. Previous methods [32, 14] are too slow. Our post classification
is also insufficient, because when high recall is expected, a cascade detector
would return too many false positives and the SVM classifier would be slow too
tion of facial points) is helpful to distinguish faces/non-faces. In Section 2,we experimentally verified that a simple SVM classifier using facial point basedfeatures in the post processing can significantly improve the accuracy of a ViolaJonesdetector. This finding is not surprising because face alignment finds correspondingparts between faces, makes them directly comparable, and simplifiesthe face/non-face classification problem. Similar observations have also beenmade in [32, 14]. While the observation is clearly useful, the real problem is howto use it effectively. Previous methods [32, 14] are too slow. Our post classificationis also insufficient, because when high recall is expected, a cascade detectorwould return too many false positives and the SVM classifier would be slow too
การแปล กรุณารอสักครู่..
