Implementation and performance of fast parallel multi-baseline stereo vision

Abstract

A fast implementation multi-baseline stereo vision on a parallel computer is described. For three 240 /spl times/ 256 images, the algorithm runs in 64 ms on 64 iWarp processors, exceeding 15 Hz frame rate. This is a speedup of 51 over an implementation on a SPARC II and represents the fastest correlation-based stereo vision system reported. Implementing this algorithm this efficiently required careful tradeoffs in algorithm design, and particularly, in the implementation of the basic communication operations. A building block approach is described for achieving best efficiency in communication; the basic operations that the parallel computer can do maximum speed are identified, and then these primitives are used to construct the communications functions needed by the algorithm.