Face Identi cation in Multimedia Archives

In this work we will present and analyse a methodology for identifying frontal-faces of characters in multimedia content. is problem is challenging due to the fact that there are a number of uncontrolled factors such as illumination, pose and facial expressions. A cascade framework of four components was implemented which (a) detects and tracks faces in a video, (b) performs face recognition by comparing the faces extracted from the video to a database of faces and (c) associates an identity to the extracted faces based on the results of the face recogniser. is framework has been evaluated on two documentary and was test both in a supervised and unsupervised environment. It was demonstrated that the problem of inaccurate training images severely affect the performance of unsupervised face identi cation systems when compared with the results achieved in the supervised setting.