xhwang/joint_image_restoration

joint_image_restoration's Issues

s value when grad_g is very small

Thanks for sharing your implementation!

In your code, you set s to be 0 when grad_g is small.
But in the original paper link, they didn't say it. Instead, according to the paper, eq. (4) and (6),
If 0 < grad_I <= epsilon and grad_g < epsilon (small values), s will be in (0, 1]
If grad_I > epsilon and grad_g < epsilon (small values), s will be in (1, 1/epsilon]

Why did you set s to be 0 when grad_g is small which is different from the paper's equation (6)? Thanks!

Inconsistency with original paper

Hello,

Thank you for the source code that you shared. I compared the results of your code with those of the binary file published by the authors of the original paper. After 6 iterations the result of your implementation is almost similar to that of the original paper,

But when we I run both methods for 700 iterations then the results are different,

As far as I looked at your code, the only difference with the original method is that you are not preconditioning the conjugate gradient solver. This must only affect the convergence rate that shouldn't matter for larger iteration counts. I was wondering if you have any idea what could be the potential issue. More specifically, I am wondering why the highlights are gone after many iterations and the image looks flat? And do you have any intuition on how the scalemap would preserve the highlights?