WinstonH.-M.HsuandShih-FuChang.PDFVIP

GENERATIVE, DISCRIMINATIVE, AND ENSEMBLE LEARNING ON MULTI-MODAL PERCEPTUAL FUSION TOWARD NEWS VIDEO STORY SEGMENTATION Winston H.-M. Hsu and Shih-Fu Chang Dept. of Electrical Engineering, Columbia University, New York {winston, sfchang}@ ABSTRACT to aggregate weak classi?ers [4, 5, 6] ef?ciently and effectively. In addition, we also develop a novel approach, called BoostME, News video story segmentation is a critical task for automatic video which combines boosting and ME; the ME classi?er is used in indexing and summarization. Our prior work has demonstrated each boosting iteration with associated con?dence scores while the promising performance by using a generative model, called Maxi- parameters and features of ME are estimated using the re-weighted mum Entropy (ME), which models the posterior probability given training data after each boosting iteration. the multi-modal perceptual features near the candidate points. In Another objective of our study is to obtain a better understand- this paper, we investigate alternative statistical approaches based ing of the problem by categorizing different production styles of on discriminative models, i.e. Support Vector Machine (SVM), video stories, analyzing the sources of classi?cation errors over and Ensemble Learning, i.e. Boosting. In addition, we develop different categories, and assessing the strengths and weaknesses of a novel approach, called BoostME, which uses the ME classi- each statistical model against different types of stories. Our ?nd- ?ers and the associated con?dence scores in each boosting iter- ings indicate that SVM-based and ME-based approaches outper- ation. We evaluated the