Occam algorithms for computing visual motion

By drawing an analogy with machine learning, the author proposes to define visual motion as a predictor that can accurately predict future frames. Under this new definition, visual motion can be specified by a collection of image patches, each moving in a simple motion. An implementation with rectangular patches determined recursively by a binary decision tree is described. Experimental results on real video sequences verify the algorithm assumptions and show that motion in typical sequences can be accurately described in terms of a few parameters.<<ETX>>