摘要

We propose a novel interactive method to generate spatial-temporal variations of human movement. Given a few examples of human movement with a particular type, user can synthesize a large number of spatialtemporal variations which are logically similar to the original motion examples by interactive controlling. Firstly, we learn a motion semantic space which can reflect the intrinsic spatial features of human motion by using a nonlinear manifold learning method ISOMAP to synthesize spatial variations. Then the temporal variations are generated by detecting and altering the movement phases. The proposed pipeline is applied to a variety of human movements. The whole process can be interactively controlled by users, which makes our method practical. The experiments show that our technique can achieve wonderful results and synthesize new motion variations efficiently.

全文