A layered stereo matching algorithm using image segmentation and global visibility constraints

TL;DR: Qualitative and quantitative results obtained for benchmark image pairs show that the proposed algorithm outperforms most state-of-the-art matching algorithms currently listed on the Middlebury stereo evaluation website.

Abstract: This work describes a stereo algorithm that takes advantage of image segmentation, assuming that disparity varies smoothly inside a segment of homogeneous colour and depth discontinuities coincide with segment borders. Image segmentation allows our method to generate correct disparity estimates in large untextured regions and precisely localize depth boundaries. The disparity inside a segment is represented by a planar equation. To derive the plane model, an initial disparity map is generated. We use a window-based approach that exploits the results of segmentation. The size of the match window is chosen adaptively. A segment's planar model is then derived by robust least squared error fitting using the initial disparity map. In a layer extraction step, disparity segments that are found to be similar according to a plane dissimilarity measurement are combined to form a single robust layer. We apply a modified mean-shift algorithm to extract clusters of similar disparity segments. Segments of the same cluster build a layer, the plane parameters of which are computed from its spatial extent using the initial disparity map. We then optimize the assignment of segments to layers using a global cost function. The quality of the disparity map is measured by warping the reference image to the second view and comparing it with the real image. Z-buffering enforces visibility and allows the explicit detection of occlusions. The cost function measures the colour dissimilarity between the warped and real views, and penalizes occlusions and neighbouring segments that are assigned to different layers. Since the problem of finding the assignment of segments to layers that minimizes this cost function is N ⁢ P -complete, an efficient greedy algorithm is applied to find a local minimum. Layer extraction and assignment are alternately applied. Qualitative and quantitative results obtained for benchmark image pairs show that the proposed algorithm outperforms most state-of-the-art matching algorithms currently listed on the Middlebury stereo evaluation website. The technique achieves particularly good results in areas with depth discontinuities and related occlusions, where missing stereo information is substituted from surrounding regions. Furthermore, we apply the algorithm to a self-recorded image set and show 3D visualizations of the derived results.