Evaluation of Mean, Gaussian and S&G aggregation windows in stereo correspondence under presence of noise

Few topics in image processing have been as extensively studied as stereo correspondence, these algorithms can be divided into two categories, local and global, depending on how the processing is done in the image. A stereo correspondence algorithm is called local if operate on sections of the images and global this treatment is performed on the entire images. In local algorithms specifically, this aggregation window is used for smoothing volume pairing cost, so that a better match is performed in presence of fronto-parallel regions. This article presents a comparison between Mean, Gaussian and Savitzky-Golay aggregation windows in local algorithms, analyzing the noise in test images and how the selection of the aggregation window affects the performance of the stereo matching algorithm.