Likelihood-constrained coupled space learning for motion synthesis

Abstract

When considering the human skeleton as a hierarchical structure (i.e. an upside-down tree structure), human motion can be treated as the rotation movements of joints around their parent joints, such that a motion sequence can be represented by the joint angle changing processes (referred to as motion patterns) of the joints between its start and end frames. With this motion representation mode, we convert motion synthesis to a motion pattern generation task, given a start-end frame pair. Conforming to such a motion synthesis scheme, we propose a coupled space learning model within which we enforce the sparse representation of the start-end frame pair on the motion frame dictionary to reconstruct its corresponding motion patterns well with a motion pattern dictionary. In addition, as motion naturalness is an important evaluation metric of motions, we propose to use the likelihood of a motion with a pre-estimated motion probability distribution to measure its naturalness and add the motion naturalness measurement into the coupled space learning model to constrain the generated motion patterns to reconstruct highly natural motions. The superior performances of both accuracy and naturalness of the synthesized motions demonstrate the effectiveness of the coupled space learning with the motion likelihood constraint.