Correspondence matching based on higher order statistics for Multi-view plus depth video sequences

This work addresses the problem of correspondence matching in multiview video sequences when co-acquired depth maps are available, as in the novel Multiview Video plus Depth (MVD) format. For the purpose of activity-based correspondence matching, we exploit the view depth information, allowing a thorough geometrical analysis of the video scene, and the statistical analysis of the inter-frame differences. In this paper we outline a correspondence matching procedure exploiting mutual information between active areas in different sequences. For activity detection we make use of the depth information and the estimated Higher Order Statistics of the inter-frame differences, which are resilient to luminance variations. The procedure encompasses two steps, namely: i) a selection of points candidate for matching search, and ii) the search of the most similar points; this procedure does not require any preliminar activity detection or segmentation stage.