The Structure of Occlusion in Fourier Space

A fundamental problem in processing sequences of images is the computation of optic ow, an approximation of the motion eld which is the projection of velocities of 3-d surface points onto the imaging plane of a visual sensor. A wide variety of algorithms for computing optical ow has been reported in the literature. However, it is not until recently that multiple motion paradigms were seriously studied. Such approaches are necessary to encompass 2-d image events such as occlusion and translucency which violate the classical single motion hypothesis. While current approaches to this problem constitute valuable contributions, they often lack the ability to determine the nature of the image events giving rise to multiple motions: Occlusion may not be diieren-tiated from translucency and it remains diicult to explicitly identify the motions associated with both the occluding and occluded surfaces. We demonstrate , under a set of reasonable assumptions, that such distinctions can be made through a Fourier analysis of these image events. We also show that translucency may be handled as a special case of occlusion.