Cost Aggregation Strategy with Bilateral Filter Based on Multi-scale Nonlinear Structure Tensor

Abstract

This paper proposed a novel cost aggregation method with modified bilateral filter. By smoothing each element of the structure tensor considering both the spatial and gradient distances of neighboring pixels, the nonlinear structure tensor for each pixel is constructed. We adopt the Log-Euclidean calculus as tensor dissimilarity function to compute the structure tensor distance of two considering pixels. Then the multi-scale value is computed by summing of the tensor distances in each scale. So a new weight basing on multi-scale structure tensor distance is set up and included in bilateral filter for cost aggregation. By adding the new weight in cost aggregation, more pixels similar with central pixel could be aggregated in a support window and the final disparity map could be more accurate. Experimental results have confirmed the effectiveness of our proposed method.