Reconstruction with Guided PatchMatch Stereo

Stereo matching is an operation that calculates a dense set of correspondences from a pair of images engineered to conform to canonical epipolar geometry. Stereo matching is a key component in many 3D reconstruction systems, thus it is a heavily studied area in the field of computer vision. Stereo matching, in the general case, is a difficult problem that remains unsolved to a satisfactory level for most scientists in the field. However, due to the large amount of attention given to this problem, there exists many algorithms, each with their own advantages and problems. This work explores the notion of combining the outputs of two well known stereo matching algorithms (symmetric dynamic programming stereo and patch match stereo) using a statistical framework, with the hope that the strategy will preserve the advantages of both algorithms while overcoming some of the weaknesses. Experiments were performed both on images from the Middlebury data set and ones captured for this study using a stereo camera. The results indicate that the proposed strategy is promising.