A novel modular framework for stereo vision

Though the theory of stereo vision is easy to grasp, the practice of computational stereo vision is much more difficult. Especially difficult is the general two-frame stereo vision problem. Many authors have published algorithms which, to a greater or lesser degrees, provide solutions to the two-frame stereo vision problem. The more recent algorithms mostly cite a landmark paper which presents a taxonomy of stereo vision consisting of some steps generally performed by stereo vision algorithms. In this paper, we propose a novel framework - the Modular Stereo Vision Framework - which describes general stereo vision problems in a more general manner than the preceding taxonomy. The central component of this framework is the cost volume, with the framework itself consisting of four stages, with each stage consisting of a number of modules. The mechanics of classifying existing and future methods as belonging to specific modules within this framework are described. Also, a discussion on the merits of the Modular Stereo Vision Framework and the preceding taxonomy is provided, focusing on the generality of the framework to current stereo algorithms and its utility for both developmental and classification purposes.