Specifying heterogeneous suites for vision tasks

To date the vast majority of parallel processing research has been restricted to the use of a single, rapid, homogeneous architecture. The primary reasons for this restriction are the uniformity of computation within the selected application and the availability of machines on which to perform the research. But, restriction one is artificial in that a multitude of applications exist that require nonuniform processing. The computer vision community provides an abundance of nice examples. Furthermore, with the current onslaught of building block components and architectures, restriction two is slowly being pushed aside. The authors describe an approach to analyzing an application with the goal of specifying an appropriate parallel processor architecture that will facilitate efficient application development (hardware and software), execution, and maintenance. Various typical computer vision tasks to illustrate the techniques are also discussed.