Linear Complexity Stereo Matching Based on Region Indexing

This paper presents a linear complexity mehtod for real-time stereo matching, in which the processing time is only dependent on the image resolution. Regions along each epipolar line are indexed to produce the disparity map, instead of searching for the best match. Current local methods have non-linear complexity, as they all rely on searching through a correlation space. The present method is limited to a parallel camera setup, because all disparities must occur in the same direction. A continuity constraint is applied in order to remove false matches. The resulting map is semi-dense, but disparities are well distributed. Experimental results on standard datasets reach around 90% of accuracy using the same parameters in all tests.