Classifiers for Motion

In this paper, we present a supervised learning based approach for sub-pixel motion estimation. The novelty of this work is the learning based method itself which tries to learn the shifts from a large training database. Integer pixel shift is sub-divided and discretized to levels in both the horizontal and vertical direction. We pose the problem of motion estimation in a polar coordinate system. Shift estimation in the x and y direction has been posed as a problem of estimating r and thetas. The ordinal property of r has been used, and consequently, we employ a ranking based approach for estimating r. For thetas estimation we employ multi-class classification techniques. We demonstrate how very simplistic features can be used to differentiate between different sub-pixel shifts