Characterization of unstructured video

In this work, we examine video retrieval from a synthesis perspective in co-operation with the more common analysis perspective. Specifically, we target our algorithms for one particular domain—unstructured video material. The goal is to make this unstructured video available for manipulation in interesting ways. I.e, take video that may have been shot with no specific intent and use it in different settings. For example, we build a set of interfaces that will enable taking a collection of home videos and making Christmas cards, Refrigerator magnets, family dramas etc out of them. The work is divided into three parts. First, we study features and models for characterization of video. Examples are VideoBook with its extensions and Hidden Markov Models for video analysis. Secondly, we examine clustering as an approach for characterization of unstructured video. Clustering alleviates some of the common problems with “query-by-example” and presents groupings that rely on the user's abilities to make relevant connections. The clustering techniques we employ operate in the probability density space. One of our goals is to employ these techniques with sophisticated models such as Bayesian Networks and HMMs, which give similar descriptions. The clustering techniques we employ are shown to be optimal in an information theoretic and Gibbs Free Energy sense. Finally, we present a set of interfaces that use these features and groupings to enable browsing and editing of unstructured video content.* (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.) *This research was supported by the Digital Life consortium.