I’m surprised that they put toy example, imagenet, and RL results in the body of the paper, but object detection results are relegated to the appendix. Is their object detection result significant? If so, why not give more attention to it?