Information-theoretic co-clustering for video shot categorization

Automatic categorization of video shots is very useful in video content analysis applications, such as structure parsing and semantic event extraction. In previous works, various low-level features, including color, texture, motion, have been used to describe the video shots. Based on these features, the similarities between shots are measured for further categorization. However, in the similarity measure, most current works treat all these feature dimensions independently, and seldom consider the potential correlations between different kinds of features. In order to explore the relationships between different features and provide a more accurate similarity measure for video shot categorization, in this paper, authors formulate the problem of unsupervised shot clustering in the scheme of information-theoretic co-clustering. In this scheme, a two-way clustering is performed to group video shots and video features simultaneously. In addition, Bayesian information criterion is employed to automatically estimate the number of clusters for both the video shots and the descriptive features. Evaluations on 1374 shots extracted from a-round 4-hour sports video shows very encouraging results in comparison with the traditional one-way clustering algorithm.