Stereo Disparity through Cost Aggregation with Guided Filter

Estimating the depth, or equivalently the disparity, of a stereo scene is a challenging problem in computer vision. The method proposed by Rhemann et al. in 2011 is based on a filtering of the cost volume, which gives for each pixel and for each hypothesized disparity a cost derived from pixel-by-pixel comparison. The filtering is performed by the guided filter proposed by He et al. in 2010. It computes a weighted local average of the costs. The weights are such that similar pixels tend to have similar costs. Eventually, a winner-take-all strategy selects the disparity with the minimal cost for each pixel. Non-consistent labels according to left-right consistency are rejected; a densification step can then be launched to fill the disparity map. The method can be used to solve other labeling problems (optical flow, segmentation) but this article focuses on the stereo matching problem. Source Code A software written in C++ is available on the IPOL web page of this article1, which is the code used in the online demo. This gives similar results to the original authors’ Matlab implementation2. The program needs several parameters (see Section 4 for more detailed explanations). By default they are tuned as suggested in the original article, but one can adapt them to get better results. Supplementary Material In the demo, an optional rectification step can be launched before running the algorithm. The source code for this preprocessing step (not reviewed) can be found at the IPOL web page of this article3.