Achieving efficient dense matching for uncalibrated images

This paper presents a new method to achieve fast dense matching in a pair of uncalibrated images. Classical area-based dense matching methods suffer from the high computational time resulting from intensive correlation calculations during the search/selection process. In contrast to conventional methods that are based on similarity and correlation techniques, this method is based on enforcing known geometric constraints and uses correlations only on a very small number of points. In particular, this paper proposes a hybrid matching technique that segment the image into two sets: the edge and the nonedge regions. For the edge regions, where discontinuities usually occur, the correlation-based classical matching method is used whereas, for nonedge regions, a segment mapping is used to achieve a correlation-free pixel matching. This segment mapping implicitly enforces all the four well known constraints in stereo matching: epipolar, continuity, uniqueness and, order constraints. The experiments on real images validated our method and showed drastic CPU-time reduction compared to classical methods.