A stochastic graph-based technique for grouping of inhomogeneous image primitives

Two main motivations are at the basis of the current interest of computer vision researchers in grouping methods: psychophysical evidence about the presence of pre-attentive mechanisms in human vision and expected reduction in computational complexity of recognition tasks. In this paper, a new probabilistic approach to grouping is proposed which is based on the representation of descriptive primitives (DPs) of different kind as sets of random variables associated with nodes of a relational graph. Grouping is modelled as the operation of assigning integer values to one among the variable of a graph node, i.e. as a labeling process. The set of random variables is described as a Markov random field with a multiple neighbourhood system. Each neighbourhood system is based on a different geometrical relation between nodes. The energy function of the field can be considered as a computational expression for some Gestalt laws which have been suggested by several psychologists as basic perceptual criteria. Two different shape descriptive primitives (i.e., circular arcs and straight segments) are here used to show the feasibility of the approach for a specific application which consists in the crowding evaluation and characterization of a surveilled environment