Gaussian scale-space dense disparity estimation with anisotropic disparity-field diffusion

We present a new reliable dense disparity estimation algorithm which employs Gaussian scale-space with anisotropic disparity-field diffusion. This algorithm estimates edge-preserving dense disparity vectors using a diffusive method on iteratively Gaussian-filtered images with a scale, i.e. the Gaussian scale-space. While a Gaussian filter kernel generates a coarser resolution from stereo image pairs, only strong and meaningful boundaries are adoptively selected on the resolution of the filtered images. Then, coarse global disparity vectors are initialized using the boundary constraint. The per-pixel disparity vectors are iteratively obtained by the local adjustment of the global disparity vectors using an energy-minimization framework. The proposed algorithm preserves the boundaries while inner regions are smoothed using anisotropic disparity-field diffusion. In this work, the Gaussian scale-space efficiently avoids illegal matching on a large baseline by the restriction of the range. Moreover, it prevents the computation from iterating into local minima of ill-posed diffusion on large gradient areas e.g. shadow and texture region, etc. The experimental results prove the excellent localization performance preserving the disparity discontinuity of each object.