Msc Thesis
Diploma thesis Face detection with Waldboost algorithm
Zdenˇk K´lal e a zdenek.kalal@gmail.com January 18, 2007
Acknowledgement
I would like to thank my supervisor, doc. Dr. Ing. Jiˇ´ Matas, for the guidance rı and advice he has provided throughout the research. I would also like to thank Jan ˇ Sochman for helping me especially in the technical issues.
Abstract
Face detection algorithms based on the work of Viola and Jones [11] train the classifier by processing training examples of face and non-face patterns. A general effort is to process a large number of training examples and hence describe the problem accurately. Current approaches are based on the assumption that non-face patterns can be easily obtained contrary to faces. While the faces remain in the training the whole time, the easy non-face patterns are stepwise replaced with more difficult ones. In the case when a large number of faces is at disposal, such conventional techniques are not able to process the whole positive training set effectively. The main contribution of this thesis is the improvement of the Waldboost algorithm [12] by which the positive training patterns are treated in a similar way as non-face patterns, called symmetric bootstrap. In the symmetric bootstrap, both the easy positive- and easy negative- patterns are stepwise replaced with more difficult ones and hence bigger space of faces is explored in the training. In order to show the effect of symmetric bootstrap, sufficiently large pool of positive training patterns had to be found and the non-face distribution appropriately modeled. A new bounding box alignment, which does not require any feature points, was found. This alignment enables us to easily collect training data from different sources (e.g. on-line face detectors) and simplifies the training task, which leads to higher detection rates of the classifier. The space of the positive training patterns was further enlarged by the technique, which is able to generate new faces
Zdenˇk K´lal e a zdenek.kalal@gmail.com January 18, 2007
Acknowledgement
I would like to thank my supervisor, doc. Dr. Ing. Jiˇ´ Matas, for the guidance rı and advice he has provided throughout the research. I would also like to thank Jan ˇ Sochman for helping me especially in the technical issues.
Abstract
Face detection algorithms based on the work of Viola and Jones [11] train the classifier by processing training examples of face and non-face patterns. A general effort is to process a large number of training examples and hence describe the problem accurately. Current approaches are based on the assumption that non-face patterns can be easily obtained contrary to faces. While the faces remain in the training the whole time, the easy non-face patterns are stepwise replaced with more difficult ones. In the case when a large number of faces is at disposal, such conventional techniques are not able to process the whole positive training set effectively. The main contribution of this thesis is the improvement of the Waldboost algorithm [12] by which the positive training patterns are treated in a similar way as non-face patterns, called symmetric bootstrap. In the symmetric bootstrap, both the easy positive- and easy negative- patterns are stepwise replaced with more difficult ones and hence bigger space of faces is explored in the training. In order to show the effect of symmetric bootstrap, sufficiently large pool of positive training patterns had to be found and the non-face distribution appropriately modeled. A new bounding box alignment, which does not require any feature points, was found. This alignment enables us to easily collect training data from different sources (e.g. on-line face detectors) and simplifies the training task, which leads to higher detection rates of the classifier. The space of the positive training patterns was further enlarged by the technique, which is able to generate new faces