Compared with EDVR or DUF?

[youyousir]

Really an exciting work! How does D3D perform under BD degradation? Have you ever compared D3Dnet with EDVR or DUF? I wonder the result

[XinyiYing]

Thanks for this comment. We have compared our D3Dnet with both EDVR and DUF-VSR methods. The comparative results between D3Dnet and EDVR are listed in Tables 6 and 7. Note that, DUF-VSR is trained on the BD degradation. For fair comparison, we retrained our D3Dnet under BD degradation. That is, HR images are first smoothed by Gaussian kernel with a standard deviation σ = 1.6 and then subsampled by selecting every 4th pixel to generate LR inputs. The training settings were kept identical to our paper. The comparative results between D3Dnet and DUF-VSR are listed in Tables 8 and 9.

It is demonstrated in Tables 6 and 8 that EDVR and DUF-VSR outperform D3Dnet in terms of PSNR and SSIM. However, as illustrated in Tables 7 and 9, the number of parameters and FLOPs of D3Dnet are 12.5% and 17.0% of those of EDVR and 44.5% and 25.4% of those of DUF-VSR. Considering the large gap of computation complexity and efficiency, we decided not to compare our D3Dnet with EDVR and DUF-VSR in the main body of our paper. The following expression is employed in our paper for clarification.

Note that, EDVR [15] and DUF-VSR [30] are not included in our comparison due to the large gap of computational cost. Specifically, the #Params/FLOPs of D3Dnet are 12.5%/17.0% of those of EDVR [15] and 44.5%/25.4% of those of DUF-VSR [30]. (Section III(C) Paragraph 1)