Estimation of image feature reliability for an interactive video segmentation scheme

We present a two-stage approach to the estimation of the reliability of different image features in the framework of a complete video segmentation scheme. The clustering method adopted in the segmentation scheme (fuzzy C-means) requires that a distance is evaluated between points in a multidimensional feature space. The analysis presented aims at establishing an appropriate weighting strategy for the different image features used in the segmentation, such as spatial information, motion, texture etc. The a priori reliability is based on an analysis of the characteristics of the frames, and influences the relative importance attributed to motion information with respect to the other features. The a posteriori reliability analysis is based on the results of the earlier stages of the segmentation process itself, and aims at identifying the features that best characterize each region and thus improve the results of the successive stages of the algorithm