Overlapping in compact pyramids

Hierarchical methods are well established in computer vision. The idea of processing images at various level of resolution is motivated by many factors, such as decreased computational cost, feasibility of divide-and-conquer methods, interaction of region based and point based operations. As a consequence, hierarchical data structures, broadly defined “pyramids”, have been introduced to experiment such ideas; among the various definitions of these data structures, overlapped pyramids have gained increasing interest especially for segmentation purposes. Meanwhile, the advantages of multi-resolution techniques have lead to the proposal of augmenting mesh-connected massively parallel architectures with pyramid like topology; a few such machines, designed to operate in SIMD or multi-SIMD mode, are in advanced stage of design or realization. It seems therefore interesting to study the effectiveness of pyramid architectures in some multiresolution algorithms, especially when the match between architecture and data structure is only partial; such is the case when machines with fixed connections among neighboring processing elements are used to run algorithms which exploit overlapped pyramids with dynamic links among adjacent levels.