A Parallel Implementation of a Correspondence-Finder for Uncalibrated Stereo Image Pairs

We report on our experience with parallelizing a computer vision algorithm. The algorithm employs low-level image processing techniques which are relatively easy to parallelize and intermediate-level computer vision techniques which lack the regularity and locality of image processing algorithms. The application is an excellent candidate for use as a benchmark. We implement two parallel versions of this algorithm; the second one based on our experience with the first version. We program the parallel implementation in the Single Program, Multiple Data (SPMD) model using the MPI message passing interface. We evaluate our implementation on a four node IBM SP. Our results show excellent speedup numbers for the image processing portion and good speedups for most of the application. However, part of the application is inherently sequential. Our second parallel implementation is not only more efficient than the first, but it also has better speedup numbers. In addition, we suggest changes to our final parallel implementation that should improve its performance.