Image ranking in video sequences using pairwise image comparisons and temporal smoothing

The ability to predict the importance of an image is highly desirable in computer vision. This work introduces an image ranking scheme suitable for use in video or image sequences. Pairwise image comparisons are used to determine image `interest' values within a standard Bayesian ranking framework, and a Rauch-Tung-Striebel smoother is used to improve these interest scores. Results show that the training data requirements typically associated with pairwise ranking systems are dramatically reduced by incorporating temporal smoothness constraints. Experiments on a coastal image dataset show that smoothed pairwise ranking can provide ranking results equivalent to standard pairwise ranking with less than half the training data.