On the Benefit of Adding an Adversarial Loss to Depth Prediction