Parallel algorithms to a parallel hardware: Designing vision algorithms for a GPU

A GPU becomes an affordable solution for accelerating a slow process on a commercial system. The most of achievements using it for non-rendering problems are the exact re-implementation of existing algorithms designed for a serial CPU. We study about conditions of a good parallel algorithm, and show that it is possible to design an algorithm targeted to a parallel hardware, though it may be useless on a CPU. The optical flow estimation problem is investigated to show the possibility. In some time-critical applications, it is more important to get results in a limited time than to improve the results. We focus on designing optical flow approximation algorithms tailored for a GPU to get a reasonable result as fast as possible by reformulating the problem as change detection with hypothesis generation using features tracked in advance. Two parallel algorithms are proposed: direct interpolation and testing multiple hypotheses. We discuss implementation issues in the CUDA framework. Both methods are running on a GPU in a near video rate providing reasonable results for the time-critical applications. These GPU-tailored algorithms become useful by running about 240 times faster than the equivalent serial implementations which are too slow to be useful in practice.