Gradual Transition Detection Using EM Curve Fitting

Abstract

Videos are made up of shots, which are connected by variant shot transitions can be categorized into hard cuts and gradual transitions. And gradual transitions can be further classified into dissolve, fade in, fade out, wipe. Shot boundary detection is the first step for video content analysis, indexing and classification. In this paper, we propose a novel approach for gradual transition detection. Our schema is to estimate the peaks on the frame to frame difference curve by EM (Expectation Maximum) curve fitting. Each peak contour is approximated by a mixture of Gaussian and uniform distributions. The weight of uniform component, the average height and the relative height of the peak are used as input features for the decision tree classifier to discriminate gradual transitions from cuts and miscellaneous. Finally, we present a framework for shot boundary detection involving camera motion detection and the combination of cut and gradual transition detection results. The advantage of our method is that it can detect all kinds of gradual transition types, such as dissolve, wipe or special effects, due to the flexibility of EM curve fitting.