Exploiting data parallelism in vision on the Connection Machine system

Abstract

The authors show how the data parallel programming model can be applied to high level vision tasks needed for object recognition. The architecture and programming model of the Connection Machine System are reviewed. Utilities for representing and manipulating sets of data, the primary representation outside of the image plane, are described using communications primitives, especially segmented scans. Several algorithms for matching and evidence accumulation, which are constructed from the utilities, are compared. The techniques emphasize the use of sorting and sparse representations of space in order to limit the combinatorial processing requirements of high-level vision.<>