Stereo Vision WITHOUT using the Smoothness Constraint

R. Chung and J. Su (PRC)


Stereo vision, Piecewise-linear approximation, Homography


Almost all existing stereo vision systems make use of the smoothness constraint to regularize the originally ill posed correspondence problem. However, the smoothness constraint tends to smooth out the occlusion boundaries which are important information about the imaged scene. We present, for any scene that could well be approximated as piecewise-linear, a stereo correspondence scheme that makes no use of the smoothness constraint in the conventional sense. The scheme, an iterative one, is based upon the interplay of two processes: (1) scene partitioning into local planar patches, and (2) inference of stereo correspondences through the use of an image-to-image mapping named homography for each of the planar patches. Experimental results with real image data are shown to illustrate the performance of the scheme.

Important Links:

Go Back