Programming intermediate level vision tasks on parallel machines

Abstract

Issues related to exploitation of parallelism in the perceptual grouping process, a generally accepted paradigm for image segmentation, are explored. The segmentation process is analyzed at three levels: tasks which are performed during a grouping process, computational requirements of algorithms that implement these tasks; and representation of data which are produced/consumed by these algorithms. Based on this analysis, the computational characteristics and requirements of these tasks for parallel implementations are identified and a programming paradigm is suggested for their parallel implementation. It consists of a data representation scheme and a set of primitive operations which have been suggested as basis of languages for data parallel programming. Pseudocode is presented for practical segmentation tasks. The suggested paradigm is compared with other programming models suggested in the literature.<>