Automatic Classification of Field of View in Video

Automatic systems are needed for audiovisual databases to efficiently index, browse, summarize and retrieve, because the amount of stored data is increasing tremendously. Historically film production techniques, have developed, in part, to convey e.g. meaning or atmosphere to the viewer. By studying these techniques, established guidelines for conveying meaning may be incorporated into automated tools for video analysis. In the current paper we present an approach in this area to classify different shot types, such as long shots, medium shots and close ups, which are important elements of video production. Based on a set of features calculated from the audiovisual content (e.g. presence of camera motion and size of detected faces), a Bayesian classifier distinguishes between six different shot types. The performance of this novel generic field of view classifier in terms of precision and recall is promising