Depth estimation via parallel coevolution of disparity functions for area-based stereo
Panos Liatsis, John Yannis Goulermas · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2001
12 A novel system for depth estimation is proposed with the use of Symbiotic Genetic Algorithms for the continuous problem of disparity surface approximation. The approach is based on the decomposition of the entire surface to very small non- overlapping patches described by low order bivariate polynomials and the use of symbiotic optimization to enforce smoothness at the boundaries of these patches, so that the entire surface can be approximated in a smooth piecewise fashion by functionals of local support. Such optimization is amenable to a massive parallel implementation, since each patch is optimized by a different execution unit and each unit communicates through its cost function only with its four-connected neighbors. The method makes use of various existing crossover and mutation schemes for real-valued chromosome representations and a new problem-specific mechanism for generating and hybridizing the initial populations. The proposed multi-objective cost function enforces photometric similarity and smoothness between the patch boundaries at a local scale, which in the long term give rise to a globally smooth disparity surface.