holmes-alan / hbpn Goto Github PK

Hi, I recently noticed your repository and read your paper. Nice try!

Meanwhile, I have some minor questions about the results:
a) The PSNR results are much higher than SOTA on Urban100, but relatively less impressive on others. Are there any particular reasons?
b) The SOTA results on benchmark datasets seem to be extremely inconsistent with HBPN RealSR (NTIRE2019) challenge rank. However, some NTIRE19 leading methods still report SOTA results on popular datasets. Which means better RealSR methods lead to higher PSNR on benchmark datasets, conversely, it cant be held?

Thanks in advance, and willing to know your opinions.

Hi,
Just want to make sure if there's typo in the paper.
Figure 3. describe the structure of UBP and DBP.
Beside the Conv op in UBP states 2Mx2NxC, which is quit confusing for me.
As far as I have understood, this Conv downsample the Deconv(x) and would be subtracted with x to construct residual map.
Shouldn't the shape be MxNx2C to fit the shape of x?
The same situation happens in DBP's Deconv.