That was great job. Can the **MotionNet** be separately trained with ground truth optical flow images?