Parallel Multiscale Stereo Matching Using Adaptive Smoothing

Abstract

We have shown a multiscale coarse-to-fine hierarchical matching of stereo pairs which uses adaptive smoothing to extract the matching primitives. The number of matching primitives at coarse scale is small, therefore reducing the number of potential matches, which in return increases the reliability of the matching results. A dense disparity can be obtained at a fine scale where the density of edgels is very high. The control strategy is very simple compared to other multiscale approaches such as the ones using Gaussian scale space, this results from the accuracy of edges detected by adaptive smoothing at different scales. The simplicity of the control strategy is especially important for low-level processing, and makes parallel implementation quite simple.