Animated movie activity characterization by image and text information fusion

In the context of animated movie characterization, we present an information fusion approach mixing very different types of data related to the activity within a movie. These data are the features extracted from images, words extracted from the synopses and expert knowledge. The difficulty of this fusion is due to the very different semantic level of these data. The aim of this work is to get a movie activity characterization in order to help the constitution of automatic summary, content based video retrieval system, etc. Two strategies are proposed : a first one aiming at giving a global description of the activity within the movie, and a second one providing a local description of activity. Tests and results are proposed on animated movies from the Annecy International Animation Film Festival.