Second - order method for occlusion relationships in motion

A new method is presented to recover the depth ordering of motion layers in a scene. We derive a measure for determining the occlusion relationship between two layers, by testing whether the support of one motion layer is moving with the predicted local velocity of another motion layer. This is used to construct a graph which represents the depth ordering of all layers. By cascading two rst-order motion models, a second-order model is deened which is sensitive to the motion of motion-deened support regions, which is sometimes called kinetic occlusion. Our algorithm makes few assumptions about the motion and layer models, and can use any motion layer model which provides global hypotheses in the form of velocity elds, and any local motion mechanism which provides the conditional probabilty of a velocity given a local image region.