Disparity Estimation based on Blind Decomposition of Images by Convolutive Kernel Compensation Technique

A recently introduced 1D blind decomposition method based on inverse correlation of a measurement matrix is examined and adapted to 2D signals. Instead of fully decomposing the image, an activity index is defined to detect areas of high activity in the input images. Activity index can also be used to detect the position of matching pixels in a stereo pair. This property is studied in more detail and applied to the problem of stereo disparity estimation. Preliminary results on artificial and real-world images show promising results, especially in the presence of strong noise and considering the highly insufficient number of available measurements