A Bayesian video modeling framework for shot segmentation and content characterization

The segmentation of video streams into their component shots is a pre-requisite for most applications involving content-based access to video libraries. In this paper, we address the segmentation problem from a probabilistic standpoint which exposes the major limitations of current solutions, and we suggest better alternatives derived from Bayesian principles. These principles lead to a framework which, by allowing the incorporation of prior knowledge about the video structure in its statistical model, leads to higher segmentation accuracy and provides a basis for content characterization which can later be used to categorize the video, retrieve it according to its content, or compare it to other instances stored in a database