Feed-forward Estimation of Optical Flow

The most common approach to generating dense optical ow elds embeds the intensity constancy constraint within a multi-resolution framework using a Laplacian pyramid. Such a framework, in which motion estimates at one scale are used to seed more accurate estimation at ner resolutions, has the potential ability to capture large motions, reducing the likelihood of aliasing. This approach, however, treats the image sequence as a series of independent processing problems where motion results from one frame do not inform the analysis of subsequent frames. In this work, this hierarchical framework is abandoned in favour of a feed-forward approach in which previous optical ow elds are fed forward (actually warped forward) to act as initial estimates in an a ne tting process in the next frame. The resulting simpli ed process architecture is not only fast but remarkably tolerant to very large pixel motions.