Questions about Downstream Face Recognition Task

[OhGaGaGaGa]

Hello authors,

We appreciate the use of downstream face recognition comparisons as a valuable metric for evaluating face restoration methods. However, neither Chimitt and Chan (authors of the paper "Simulating Anisoplanatic Turbulence by Sampling Inter-modal and Spatially Correlated Zernike Coefficients") nor you have provided the degradation code. Could you please explain the "unseen atmospheric turbulence degradation" in more detail? If possible, it would be great if you could release the degradation code or provide the degraded LFW dataset from 10000 to 40000.

Your work is impressive. Thank you for your assistance!

[LIAGM]

Hi,

Due to the internal project issue, we can not release the related code and data.
I suggest you find some related code on Prof. Chan's website to generate the turbulence degradation.
Thanks for reaching out!

[OhGaGaGaGa]

Thanks so much for your prompt and helpful responses!

We have thoroughly reviewed the Python code available on TurbulenceSim_v1 Repository provided in Prof. Chan's website and have a couple of questions we hope you could assist us with:

1. Degradation Parameters: Does the term "degradation parameters" in DAEFR refer to the "length of propagation (meters)" in the code? We want to ensure that we are accurately interpreting and applying these parameters.

2. RGB Image Processing: The provided code primarily supports processing grayscale images. We attempted to apply atmospheric turbulence degradation to each RGB channel separately; however, the resulting images exhibit red and green edges, which are inconsistent with the results in your paper's supplementary materials. Could you please provide guidance on correctly handling RGB images to achieve the expected degradation?

We are relatively new to atmospheric turbulence degradation. Any insights, recommendations, or additional resources you could provide would be valuable to us. Thank you very much for your assistance.

[LIAGM]

Hi,

Here is the answer to your questions:

1. Yes, we mainly change the "length of propagation (meters)" from 10000 to 40000 and also let the "size of the image" be the original image size.
2. We also process the RGB channel separately and do not use the "rgb2gray" function provided by the author.

I hope these responses can help you!