Estimating disparity and motion using multiresolution Fourier analysis

The paper describes a multiresolution approach to tackling the image registration problem-the identification of corresponding points or regions in two or more images of a scene derived, for example, from different viewpoints or at different times. It is a region-based approach in which disparity estimates obtained from local correlations computed via the frequency domain are incorporated within a multiresolution search procedure. As such the approach removes the link between frequency content and disparity, while retaining the efficiency of multiresolution searching. The scheme is implemented within the framework of a generalised wavelet transform, the multiresolution Fourier transform, which also provides the potential for the incorporation of more complex matching criteria based on feature information and affine transformation. The aim is to outline the main components of the basic algorithm and present results of its use in estimating the disparity between binocular views of a scene and in estimating motion from video sequences.