Occam algorithms for computing visual motion

The standard approach to computing motion relies on pixel correspondence. Computational schemes impose additional constraints, such as smoothness and continuity of the motion vector even though these are not directly related to pixel correspondence. This paper proposes an alternative to the multiple constraints approach. By drawing analogy with machine learning, motion is computed as a function that accurately predicts frames. The Occam-Razor principle suggests that among all functions that accurately predict the second frame from the first frame, the best predictor is the “simplest,” and simplicity can be rigorously defined in terms of encoding length. An implementation of a practical algorithm is described. Experiments with real video sequences verify the algorithm assumptions by showing that motion in typical sequences can be accurately described in terms of a few parameters. Our particular choice of predictors produces results that compare very favorably with other image flow algorithms in terms of accuracy and compactness. It may, however, be too constrained to enable accurate recovery of 3D motion and structure