Disparity Component Matching Revisited

In this paper, we demonstrate how the quality of disparity maps could be improved by applying windows that adapt their shape. In our approach, windows adapt to high-correlation structures in disparity space, which we call the disparity components. The shape of adaptive windows is found by projection the disparity components to the input images, which results in two non-rectangular windows. The method is straightforward and it is applicable to any stereo algorithm. It differs from other approaches in: (1) the way the disparity components adapt to structures of non-constant disparity, (2) the way the disparity components are used for re-computing the similarity values. This helps us in solving standard stereo problems such as: enforcing the inter-line continuity constraint, or to improving disparity map accuracy. To show the proposed method properties, we measure the improvement of stereo matching failure due to adaptive windows in a rigorous ground-truth evaluation experiment. We demonstrate that adaptive windows account for 15× improvement in false positive rate and 5× improvement in mismatch rate. Another quantitative comparison, based on Scharstein & Szeliski’s methodology, is performed in detail in order to confront our results with results of various existing stereo methods. Finally, the results on other standard stereo image pairs are presented.