In simple situations, both Laplacians are zero. If the viewer translates parallel to a flat object, rotates about a line perpendicular to the surface or travels orthogonally to the surface, then the second partial derivatives of both u and v
vanish (assuming perspective projection in the image formation). We will use here the square of the magnitude of the gradient as smoothness measure. Note that our approach is in contrast with that of Fennema and Thompson [S], who propose an algorithm that incorporates additional assumptions such as constant flow velocities within discrete regions of the image. Their method, based on cluster analysis, cannot deal with rotating objects, since these give rise to a continuum of flow velocities.