Parallel processing in the DARPA strategic computing vision program

Hardware, software tools, algorithms, and performance metrics that have been developed for image understanding are presented. Three commercially built examples reflecting three mature approaches considered germane to vision-single-instruction multiple-data, multiple-instruction multiple-data, and systolic processing-were chosen. They are, respectively, the Connection Machine, the Butterfly, and the Warp. A fourth approach, more specific to vision, was also selected for noncommercial implementation. This machine, the Image-Understanding Architecture, involves a heterogeneous combination of parallel processors with single-instruction multiple-data, multiple-instruction multiple-data, and other capabilities. Each site employing one of the above architectures developed a different set of tools, leading to significant cross-fertilization of ideas between the sites. Algorithms for low-level vision, shape from texture, fusing stereo and texture, surface interpolation, and robot navigation, among others, are briefly discussed. Benchmarks are described.<<ETX>>