Fusion of Multiple Cues for Video Segmentation

The segmentation of video into contiguous scenes is becoming an important problem in many applications. Since video data is a rich source of spatio-temporal information, diierent types of features can be computed in the video data. Each of these features provide a cue for the segmenta-tion of video and is usually suucient to perform an approximate segmentation of video. However, the features many times provide connicting evidence for segmentation. Further, since there is strong correlation between the diierent features, it is not easy to fuse the information from the features to make segmentation decisions. We present a method based on Bayesian Networks that model the dependence between the segmentation decision and the diierent features. This framework using Bayesian Networks is promising and provides an extensible mechanism for fusion of information.