Adaptive piecewise elastic motion estimation

Abstract

This paper proposes a novel probabilistic graphical model, called MTHMM-P, for partitioning general nonrigid motion into piecewise elastic motion, so as to achieve nonrigid motion estimation without the need of any a priori shape model. To this end, two interrelated sub-problems have to be addressed: partitioning whole motion sequence into several coherent pieces and estimating elastic motion inside each one. The proposed MTHMM-P jointly formulates these two sub-problems. By means of a competition-cooperation partition mechanism and joint inference algorithm, the MTHMM-P can automatically determine the number of pieces as well as adjust the span of each piece by adapting to the input sequence, and at the same time achieve the estimation of piecewise elastic motion. Experiments on the motion of face and body show the capability of the MTHMM-P.