|
Abstract : |
Abstract. This paper presents a method for extending upright, frontal, template-based face detection systems to efficiently handle all in-plane rotations. Detecting rotated faces is a two step procedure. First a "DeRotation " network is used to process each input window. If there is a face in the window, this network determines its angle of rotation. Based upon this estimated angle of rotation, the window is then rotated to an upright position. Second, a "Detection " network, or multiple detection networks, are used to determine whether the rotated window contained an upright face. The training methods for both networks are presented. Preliminary empirical results are also provided. 1, |
