Multi-level Motion-Informed Approach for Video Generation with Key Frames

摘要

Observing that a motion signal is decomposable into multiple levels, a video generation model which realizes this hypothesis is proposed. The model decomposes motion into a two-level signal involving a global path and local pattern. They are modeled via a latent path in the form of a composite Bezier spline along with a latent sine function respectively. In the application context, the model fills the research gap in its ability to connect an arbitrary number of input key frames smoothly. Experimental results indicate that the model improves in terms of the smoothness of the generated video. In addition, the ability of the model in separating global and local signal has been validated.