Self-Supervised Learning of Object Motion Through Adversarial Video Prediction