Novel methods for discontinuity preserving support aggregation in local stereo matching

Abstract

In this thesis we focus on local stereo matching methods, which rely on finding corresponding image pixels between the left and right image of a stereo pair by calculating aggregated color dissimilarities of pixels within suitable image patches. We propose novel adaptive support weight aggregation techniques that advance the state-of-the-art in local stereo matching in terms of both reconstruction quality and computational speed. (i) We propose an improved method for computing the adaptive support weights by imposing an additional connectivity constraint that relies on a geodesic distance transform. (ii) In order to overcome the problem of the large processing time, we present an algorithm for considerably speeding up computation times of the cost aggregation step of the adaptive support weight techniques. (iii) From the cost filtering/smoothing point of view, we introduce a stereo matching algorithm which uses the fast guided image filtering strategy for edge-preserving cost aggregation. (iv) We also propose an efficient algorithm for computing temporally consistent depth maps from video sequences. (v) To help researchers/developers selecting among different weight computation techniques, we conduct an extensive experimental evaluation of a variety of support weight computation methods suggested in the literature or proposed by ourselves. This evaluation examines the accuracy of the retrieved depth maps and the computational efficiency of a corresponding GPU-based implementation. (vi) Finally, we demonstrate that our proposed stereo algorithms that rely on cost filtering can be extended successfully to the closely related optical flow (motion) problem.