Co-segmentation of image pairs with quadratic global constraint in MRFs

This paper provides a novel method for co-segmentation, namely simultaneously segmenting multiple images with same foreground and distinct backgrounds. Our contribution primarily lies in four-folds. First, image pairs are typically captured under different imaging conditions, which makes the color distribution of desired object shift greatly, hence it brings challenges to color-based co-segmentation. Here we propose a robust regression method to minimize color variances between corresponding image regions. Secondly, although having been intensively discussed, the exact meaning of the term "co-segmentation" is rather vague and importance of image background is previously neglected, this motivate us to provide a novel, clear and comprehensive definition for co-segmentation. Thirdly, it is an involved issue that specific regions tend to be categorized as foreground, so we introduce "risk term" to differentiate colors, which has not been discussed before in the literatures to our best knowledge. Lastly and most importantly, unlike conventional linear global terms in MRFs, we propose a sum-of-squared-difference (SSD) based global constraint and deduce its equivalent quadratic form which takes into account the pairwise relations in feature space. Reasonable assumptions are made and global optimal could be efficiently obtained via alternating Graph Cuts.