Exploiting Discontinuities in Optical

Most optical ow estimation techniques have substantial diiculties dealing with ow discontinuities. Methods which simultaneously detect ow boundaries and use the detected boundaries to aid in ow estimation can produce signiicantly improved results. Current approaches to implementing these methods still have important limitations, however. We demonstrate three such problems: errors due to the mixture of image properties across boundaries, an intrinsic ambiguity in boundary location when only short sequences are considered, and diiculties insuring that the motion of a boundary aids in ow estimation for the surface to which it is attached without corrupting the ow estimates for the occluded surface on the other side. The rst problem can be xed by basing ow estimation only on image changes at edges. The second requires an analysis of longer time intervals. The third can be aided by using a boundary detection mechanism which classiies the sides of boundaries as occluding and occluded at the same time as the boundaries are detected.