Multiresolution pyramid techniques can improve the efficiency of basic vision algorithms by orders of magnitude. They will be key to developing practical vision systems to perform challenging tasks in real time. However, these efficiencies cannot be fully realized when pyramid-based algorithms are implemented on machines with conventional SIMD mesh and pipeline architectures. We describe a segmented pipeline architecture that can support pyramid processing. We illustrate this architecture with an application to image motion analysis for vehicle guidance.
Peter J. Burt