Accurate Segmentation and Estimation of Parametric Motion Fields for Object-based Video Coding using Mean Field Theory

We formulate the problem of decomposing a scene into its constituent objects as one of partitioning the current frame into objects comprising it. The motion parameter is modeled as a nonrandom but unknown quantity and the problem is posed as one of Maximum Likelihood (ML) estimation. The MRF potentials which characterize the underlying segmentation field are defined in a way that the spatio-temporal segmentation is constrained by the static image segmentation of the current frame. To compute the motion parameter vector and the segmentation simultaneously we use the Expectation Maximization (EM) algorithm. The E-step of the EM algorithm, which computes the conditional expectation of the segmentation field, now reflects interdependencies more accurately because of neighborhood interactions. We take recourse to Mean Field theory to compute the expected value of the conditional MRF. Robust M-estimation methods are used in the M- step. To allow for motions of large magnitudes image frames are represented at various scales and the EM procedure is embedded in a hierarchical coarse-to-fine framework. Our formulation results in a highly parallel algorithm that computes robust and accurate segmentations as well as motion vectors for use in low bit rate video coding. This report has been submitted as a paper to the SPIE conference on Visual Communications and Image Processing - VCIP98 to be held in San Jose, California on Jan 24- 30, 1998.