Practical considerations for executing vision algorithms on parallel arrays

It is shown that levels of performance claimed for vision algorithms can be overly optimistic if the computation time and communication time, both of which affect the schedule length of algorithms running on parallel arrays, are not quantified. Examples are given to show how the embedding technique and problem partitioning can affect communication time. When the number of computations exceeds the number of processing elements, it is necessary to share resources, but care must be taken in how this is done to minimize the schedule length. It is also shown that the scheduling technique chosen directly affects speedup