Segmentation with pairwise attraction and repulsion

We propose a method of image segmentation by integrating pairwise attraction and directional repulsion derived from local grouping and figure-ground cues. These two kinds of pairwise relationships are encoded in the real and imaginary parts of an Hermitian graph weight matrix, through which we can directly generalize the normalized cuts criterion. With bi-graph constructions, this method can be readily extended to handle nondirectional repulsion that captures dissimilarity. We demonstrate the use of repulsion in image segmentation with relative depth cues, which allows segmentation and figure-ground segregation to be computed simultaneously. As a general mechanism to represent the dual measures of attraction and repulsion, this method can also be employed to solve other constraint satisfaction and optimization problems.