This talk will address a long-standing problem in motion analysis, namely, the detection and estimation of motion in the neighborhoods of surface boundaries. Motion in these regions is discontinuous, and occlusions cause image structure to appear or disappear from one image to the next. Although these "motion boundaries" are often viewed as a source of noise for current motion estimation techniques, we can also view them as a rich source of information about the location of surface boundaries and the depth ordering of surfaces at these locations.
We propose a Bayesian framework for representing and estimating image motion in terms of multiple motion models, including both smooth motion and local motion discontinuity models. We compute the posterior probability distribution over models and model parameters, given the image data, using discrete samples and a particle filter for propagating beliefs through time. This talk will introduce the problem and describe our Bayesian approach, including our generative models, the likelihood computation, the particle filter, and a mixture model prior from which samples are drawn. I will present several experimental results on tracking motion discontinuities, and if time permits I will also show results of some related projects in the Digital Video Analysis (DiVA) group at Xerox PARC.