A generative model based approach to motion segmentation

We address the question of how to choose between different likelihood functions for motion estimation. To this end, we formulate motion estimation as a problem of Bayesian inference and compare the likelihood functions generated by various models for image formation. In contrast to alternative approaches which focus on noise in the measurement process, we propose to introduce noise on the level of the velocity, thus allowing it to vary around a given model. We show that this approach generates additional normalizations not present in previous likelihood functions. We numerically evaluate the proposed likelihood in a variational framework for segmenting the image plane into domains of piecewise constant motion. The evolution of the motion discontinuity set is implemented using the level set framework.