Grouping with Directed Relationships

Abstract

Grouping is a global partitioning process that integrates local cues distributed over the entire image. We identify four types of pairwise relationships, attraction and repulsion, each of which can be symmetric or asymmetric. We represent these relationships with two directed graphs. We generalize the normalized cuts criteria to partitioning on directed graphs. Our formulation results in Rayleigh quotients on Hermitian matrices, where the real part describes undirected relationships, with positive numbers for attraction, negative numbers for repulsion, and the imaginary part describes directed relationships. Globally optimal solutions can be obtained by eigendecomposition. The eigenvectors characterize the optimal partitioning in the complex phase plane, with phase angle separation determining the partitioning of vertices and the relative phase advance indicating the ordering of partitions. We use directed repulsion relationships to encode relative depth cues and demonstrate that our method leads to simultaneous image segmentation and depth segregation.