A causality-attentive stereo matching method for shape-preserved depth map

Abstract

The preservation of object shape and uniform-area flatness in depth maps estimated by local stereo marching approaches is critical and still challenging. Toward the achievement of such preservation in the resulted depth map, we promote the stereo-matching algorithm by exploring and creating the causal relations while realizing the algorithmic construction in this paper. By spatial attentive causality, the decision of selected anchor pixel precedes for imposing its disparity as the range constraint of disparity searching on the other pixels in every horizontal line. By contextual attentive causality, the proportional factor of local-to-global gradient density can be applied to select the size of adaptive support weight (ASW) window. For feature-causality creation, the major cost term is delivered as the model parameter to compute other cost terms for forging the rational tie-in links between cost terms. During the comparison of ASW matching costs, the possible candidates are simultaneously confirmed for substituting the suspicious primary disparity in the subsequent content-aware refinement. For inducing the spatial attentive causality, the super-pixels are commended to enhance the cost weighting and the hole filling which is an effective dual inpainting routine. By realization and deployment of the causalities, the depth maps resulted from the proposed adaptive-supporting method will be apt to acquire the preservation of object shape and the suppression of over-shooting error in flat regions. The experimental results demonstrate that the proposed method based on the causality findings can obtain low bad-pixel rate and high visualization quality, both.