question on image representation results

[jypppppp]

Dear author,

When I redo the image representation on the "lighthouse" image, the result is 25 dB which is a big difference from 43.2 dB and the time is around 3 min. I run it on "wire_occupancy.py" and changed layer numbers to 3 and neuron numbers to 300.

Is what I did right?

Thank you

Jason

[vishwa91]

Thank you for the question. Two changes to replicate the result:

1. Downsample image by 1/4 -- this is to ensure that the number of parameters in the network, and image complexity (pixels) is similar.
2. Set omega=20, and sigma=20 -- as we discussed in the paper, the set of (omega, sigma) for image representation and image denoising are different. Intuitively, for representation, we want higher frequencies overall, and hence larger omega, sigma.

Hope that helps!

[jypppppp]

Thanks for your reply and it helps a lot.

Also, can you provide the code for single image super-resolution?

Jason

[vishwa91]

Will upload SISR code within a week!

[skxgogo]

I found an interesting phenomenon that when the 1e-4 learning rate was used, the output of models can't accurately represent the color of the image. But with the 5e-3 learning rate , there is no problem. Even the small learning rate has the same psnr with large learning rate