Questions about indicators

[ct-hub]

Hello, I observed. In this paper, the index of PSNR-IR_GIR was lower than 30dB. In Zhao's article, his PSNR index is all higher than 60dB. It's confusing to me

[JorgeFCS]

Hi! Yes, that's true, the PSNR results reported by Zhao et al. are higher. However, when I implemented and tested their proposed architecture, the resulting PSNR scores were lower.

Zhao et al. train and validate the model on the TNO Dataset, doing some pretty intense data augmentation. It is possible that the difference in results is due to the database employed.

Another possibility could be that some of the missing implementation details (e.g. # of filters per layer, etc.) could be responsible for the performance difference; in the README I mention that I refer to two previous works (Ma et al., Isola et al.) to fill in these gaps.