Towards stereo from scale trees

The image trees described in this paper hierarchically organize image segments according to scale, with the coarsest scale, the scale of the image itself, as the root of the tree and the finest scales as the leaves. The segmentation algorithm used to form the nodes of the tree is the sieve, a nonlinear morphological scale-space operator. We use a conventional matching criterion, the sums of squares of differences, SSD, which is statistically meaningful when ergodicity of its matching regions is assumed. Here, simple statistical similarity measures are applied to the scale tree to simplify it into a set of relatively homogeneous regions. This simplified scale tree is then used to generate matching regions which frequently satisfy the ergodicity condition. This approach reduces the errors in the resulting disparity map particularly within sharp-edged regions with low texture-where conventional methods fail