Human beings process stereoscopic correspondence across multiple scales.However, this bio-inspiration is ignored by state-of-the-art cost aggregationmethods for dense stereo correspondence. In this paper, a generic cross-scalecost aggregation framework is proposed to allow multi-scale interaction in costaggregation. We firstly reformulate cost aggregation from a unifiedoptimization perspective and show that different cost aggregation methodsessentially differ in the choices of similarity kernels. Then, an inter-scaleregularizer is introduced into optimization and solving this new optimizationproblem leads to the proposed framework. Since the regularization term isindependent of the similarity kernel, various cost aggregation methods can beintegrated into the proposed general framework. We show that the cross-scaleframework is important as it effectively and efficiently expandsstate-of-the-art cost aggregation methods and leads to significantimprovements, when evaluated on Middlebury, KITTI and New Tsukuba datasets.
展开▼