Evaluating the covariance matrix constraints for data-driven statistical human motion reconstruction

Abstract

This paper presents the evaluation process of the character's motion reconstruction while constraints are applied to the covariance matrix of the motion prior learning process. For the evaluation process, a maximum a posteriori (MAP) framework is first generated, which receives input trajectories and reconstructs the motion of the character. Then, using various methods to constrain the covariance matrix, information that reflects certain assumptions about the motion reconstruction process is retrieved. Each of the covariance matrix constraints are evaluated by its ability to reconstruct the desired motion sequences either by using a large amount of motion data or by using a small dataset that contains only specific motions.