On-line characters identification in movies

Characters identification in video consists in assigning name labels to persons present in a video. We explore here the online labeling of faces in movie. Previous work such as demonstrated promising results on learning and classifying characters using a manually annotated learning-corpus. Some practical issues appear when applying this method on large scale movie database in permanent evolution, as the number of characters to recognize is important and continuously grows. In this paper we build on the first method extending the coverage greatly by learning appearance models of new cast members online for each new movie. In addition, we make the following contributions: (1) we propose to apply Active Appearance Models (AAM) tracking method in order to track local facial features over time and orientation changes, (2) we evaluate important parameters of the feature extraction such as position and size of local features. We report results on the movie I am a legend demonstrating the relevance of our on-line approach of the problem.