Thanks again for your work. Since you use flip and multi-scale strategy in testing, I wonder how your model performs when these test tricks are not applied. I ask this question because in real-time application sceneries, these tricks are not useful.