Experiments in estimation of independent 3D motion using EM

In this paper we address the problem of multiple 3D rigid body motion estimation from the optical flow. We use the differential epipolar constraint to measure the consistency of the local flow estimates with 3D rigid body motion and employ a probabilistic interpretation of the overall flowfield in terms of mixture models. The estimation of 3D motion parameters as well as the refinement of the initial motion segmentation is carried out using an Expectation-Maximization (EM) algorithm. The algorithm is guaranteed to improve the overall likelihood of the data. The proposed technique is a step towards estimation of 3D motion of independently moving objects in the presence of egomotion.