iGG-MBS: Iterative Guided-Gaussian Multi-Baseline Stereo Matching

This paper presents an improved dense disparity estimating technique for a collection of multi-baseline stereo (referred to as MBS in the text) images. The flow of the proposed system consists of two main frameworks: a preliminary cost calculation and initial disparity estimating framework, and an iterative cost refinement framework. The first framework implements an accurate multi-baseline stereo cost (referred to as MBSC in the text) calculation method, and a scan line optimization inspired by the Semi Global Matching (SGM) algorithm. Cost volumes of each two-view camera pair are calculated by fusing two pixel dissimilarity measures: i) weighted Census transformation and ii) sum of absolute difference color consistency term (SAD-Census). The initial disparity map between reference and the matching view with the largest baseline displacement is calculated by summing-and-interpolating SAD-Census costs of the current and all neighboring camera pairs in-between, and taking the minimum after aggregating for sixteen directions. The second framework refines the aggregated MBSC volume recursively. In each iteration, individual pair-wise disparity maps are used to warp matching views towards the reference to create binary masks that resemble overlapping differences. White locations in the mask represent incorrect correspondence matches, thus a penalty is added for costs associated with, adapting a Gaussian modulating function. This significantly reduces the selection probability of incorrect disparity minima in proceeding iterations. A Guided filter-based Rolling Guidance filter is applied to further up-vote the probability of pixels with the lowest costs, which are similar or close enough to ground truth readings. Through experimental results evaluated on the Middlebury dataset, we show that our method leads to effective and efficient multi-baseline disparity estimations.