Forums

> How to synthesize low-resolution noise images in the challenge 1

There are three questions about how to synthesize low-resolution noise images.

1. About Noise Level
In the evaluation rules, the noise level is described as ----- (gauss, m=0, s=10).
But in the past two relevant workshop papers, the noise level is described as ----- Gaussian noise (σ = 10%) is
added.
These two statements don't seem to mean the same thing.

2. Downsampling first or Adding noise first?
In the evaluation rules, it is described as ----- noised (gauss, m=0, s=10) & bicubic downsampled (adding noise first).
But in the past two relevant workshop papers, it is described as ----- A downsampling process with a scale factor of
×4 is applied. Additionally, Gaussian noise (σ = 10%) is added. (downsampling first)
These two statements don't seem to mean the same thing.

3. Published test sets
The published test set appears to have a very low noise level, we have tried various combinations: downsampling first and adding noise first, sigma is 10% in the range of [0,1] and sigma is 10 in the range of [0,255]. None of these attempts on the validation set produced a picture like the noise level of the test set. In fact, the noise level of the images generated on the validation set was higher than that of the test set, from a subjective visual point of view.

In summary, can the organizers provide codes on how to synthesize noisy low-resolution images to standardize? The code that generates the test set images is just fine because the training and test cases should be uniform.