A Novel Trilateral Filter based Adaptive Support Weight Method for Stereo Matching

The performance of local stereo matching algorithm highly depends on the support window selected, in which the cost are aggregated. A variety of cost aggregation approaches (proposed before 2008) were comprehensively analyzed in [7] and these approaches attempt to seek an optimal support window for each pixel by changing the window size, shape and center offset. The ideal optimal window should satisfy the rule that all pixels in this window lie on the same disparity with the center pixel. Recent years have witnessed a great deal of attention focused on the Adaptive Support Weight (ASW) based methods [1, 4, 6, 8], proposed firstly by Yoon and Kweon in [8]. The ASW methods assign an adaptive weight to each pixel of the support window, depending on how it is likely to lie on the same disparity with the center pixel. Essentially, the assignment of an adaptive weight amounts to changing the support window in terms of size, shape or center offset. In ASW methods, the weight function is very important, because it directly decides the support window. The weight function proposed in [8] is based on bilateral filter. Following this pioneering work, various weight function were proposed, including in particular the segmented bilateral filter weight function [6], the geodesic weight function [1], the guided filter weight function [4]. Thus, which weight function is the most accurate one? Recently, Hosni et al. [2] carried out a comprehensive comparative study to fairly evaluate various weight functions while fixing the preprocessing, matching cost function and post-processing. Their conclusion is that both bilateral filter weight function [8] and guided filter weight function [4] are the best, since bilateral filter weight function performs better on the average rank while guided filter weight function produces a lower average error. We revisit the bilateral filter weight function [8], which obeys two rules that, given a support pixel, (1) if its color is similar to the center pixel’s and (2) if it is spatially close to the center pixel, it is likely to lie on the same disparity with the center pixel. Therefore, the bilateral filter weight function consists of two parts, color similarity term and spatial proximity term, defined as: