![inpaint review inpaint review](https://2.bp.blogspot.com/-jQmzdZBb1BA/WofU4MWv5aI/AAAAAAAAAI4/V1iZ-5JknX0zOsuHsLS93xoOhgjdNgVdgCLcBGAs/s1600/Teorex-Inpaint-7-Crack-Full-Keys.png)
fall streaks) this results in an unlikely but still very realistic looking sample. For weather features whose shape is heavily determined by the large-scale flow (e.g. For most cloud features that are embedded in the large-scale flow this results in a physically plausible sample and approximates large scale flow in the opposite direction. For KaZR data, flips were performed with respect to time (not in the vertical). Many of the data augmentation schemes used in the deep learning literature were designed for use with images and, if applied to radar data, would result in weather features that are physically impossible, so we were careful here to only use augmentations that produce physically plausible samples. We moved the discussion of data augmentation to the end of Section 3.4 (“Training” Lines 271-282) and expanded it. This is a good point, and we think data augmentation deserves a more thorough treatment in the manuscript. We sincerely appreciate you taking the time to review our work and provide constructive comments! We have used your feedback to improve the paper and have responded to each comment in order below:
![inpaint review inpaint review](https://i.pcmag.com/imagery/reviews/06JQnPuKCHyQZNr0z9tV116-38.png)
Thank you to both reviewers and to the editor for your thoughtful comments.
Inpaint review how to#
I know this is addressed a bit already, but I think the outstanding qualitative skill of the CGAN (while poor quantitative skill) requires a bit more discussion of the risks, how they might be addressed for applications, and some speculation on how to leverage the relative skills of the two novel techniques. Also, it's skill has risks as mentioned above. It has the poorest performance, underperforming all the non-NN techniques on certain metrics. The CGAN's creativity in hallucinating weather patterns could pose risks if the presence of speckle or other faint, spurious data could trigger inpainting of weather patterns that do not exist.Ĥ10, 420, 455: The CGAN success should be discussed a little more. What about the performance on a sparsely speckled background? In real time radar applications, masks can be leaky due to atmospheric and instrumental variations. Is there any way to quantify whether these add skill? Are there any downsides to flipping a meteorological image, thereby including physically impossible weather patterns in the training set?ģ35: The performance on blank space is dismissed as trivial. I have a few criticisms and questions, but they are meant only to clarify certain points in the conclusions and perhaps illustrate some of the authors comments.ġ30: Flips and rotations of images are used. The paper clearly and compellingly expands computer vision inpainting into new meteorological applications. The methods are valid, and presentation well in line with the greater computer vision inpainting literature.