Spatio-temporal Segmentation Using Dominant Sets

Abstract

Pairwise data clustering techniques are gaining increasing popularity over traditional, feature-based central grouping techniques. These approaches have proven very powerful when applied to image-segmentation problems. However, they are computationally too demanding to be applied to video-segmentaton problems due to their scaling behavior with the quantity of data. On a dataset containing N examples, the number of potential comparisons scales with O(N ²), thereby rendering the approaches unfeasible for problems involving very large data sets. It is therefore of primary importance to develop strategies to reduce the number of comparisons required by subsampling the data and extending the grouping to out-of-sample points after the clustering process has taken place. In this paper we present an approach to out-of-sample clustering based on the dominant set framework [10] and apply it to video segmentation. The method is compared against two recent pairwise clustering algorithms which provide out-of-sample extensions: the Nyström method [3], and the minimal-shift embedding approach [14]. Our results show that our approach performs comparably against the competition in terms of quality of the segmentation, being, however, much faster.