Medium knowledge-based macro-segmentation of video into sequences

EEcient and reliable methods for the segmentation of the image part of a video into shots have been proposed. In actual motion picture or video documents, there can often be 500 to 1000 shots per hour. Thus, if one wants to enable quick browsing of the video contents, quick positioning in the document for interactive viewing, or if one wants to automatically construct abstracts of the document, it is necessary to nd more macroscopic time objects, for instance larger sequences constituting a narrative unit or sharing the same setting. In this paper, we present a method for obtaining automatically such a macro-segmentation in sequences. This method is based on the application of rules expressing the local (in time) clues which are given by the medium contents to enable identiication of more macroscopical changes. We describe how the results from these rules can be combined to obtain a macro-segmentation, and to extract particularly important representative images. We give arguments for the choice of such a general purpose medium-based approach compared to approaches based on modelling speciic types of contents, and present results from its automatic application to a limited sample of documents. This research was conducted under support from the french Minist ere de la Culture et de la Francophonie.