Dense Stereo Using Pivoted Dynamic Programming

Abstract This paper describes an improvement to the dynamic programming approach for dense stereo. Traditionally dense stereo algorithms proceed independently for each pair of epipolar lines, and then a further step is used to smooth the estimated disparities between the epipolar lines. This typically results in a streaky disparity map along depth discontinuities. In order to overcome this problem the information from corner and edge matching algorithms are exploited. Indeed we present a unified dynamic programming/statistical framework that allows the incorporation of any partial knowledge about disparities, such as matched features and known surfaces within the scene. The result is a fully automatic dense stereo system with a faster run time and greater accuracy than the standard dynamic programming method.