Disparity map computation on scalable computing

This paper addresses the evaluation of a new disparity map computing algorithm characterized by a novel spurious removal strategy. Using this algorithm we eliminate a high percentage of wrong values with a low performance penalty. When testing images, incorrect percentages were reduced by 65% and 85%. This algorithm has been designed for scalable architectures with massive parallel processing elements. It works line by line with low memory requirements. To evaluate the scalability of the algorithm we have implement it on two different architectures, a GPU and a CPU: The GPU was a NVIDIA GTX260 graphic card using CUDA. The CPU was an Intel Core i7 920 with 4 cores and hyperethreading (8 virtual cores). Our implementation of the algorithm on the CPU was capable of using an arbitrary number of threads. The tests presented in this paper show that the CPU performance scales from 1 to 4 threads with a factor of 0.66, while the speedup when comparing the GPU with the single-threaded CPU solution is between 38 and 47 times. From the power consumption point of view, the GPU is more than ten times more efficient than the CPU.