I see that your paper describes that the middle frame is inferred based on four frames, but there are currently only demos inferred from two frames on GitHub, how can I infer nonlinear motion based on the acceleration information of the object as mentioned in the paper?