Motion-based classification of cartoons

Abstract

This paper describes a simple high-level classification of multimedia broadcast material into cartoon non-cartoon. The input video sequences are from a broad range of material which is representative of entertainment viewing. Classification of this type of high-level video genre is difficult because of its large inter-class variation. The task is made more difficult when classification is over a small time (10's of seconds) introducing a great deal of intra-class variation. This paper presents a purely dynamic based approach for content-based classification of video sequences in the form of a new global motion measure of foreground objects. Experiments are reported on a diverse database consisting of: 8 cartoon and 20 non-cartoon sequences. Results are shown in identification error rates against time of sequence used for classification. The system produces a best identification error rate of 3% on 66 separate decisions based on 23 second sequences trained using a total of /spl sim/20 minutes of video.