Why do you use nearest method for matching the resolution of (LR, HR) due to CutBlur ?

[star4s]

I have a question about how to match he resolution of (LR, HR) due to CutBlur.

When I check the code about matching the resolution of (LR, HR) due to CutBlur,
I found using nearest.

match the resolution of (LR, HR) due to CutBlur

        if HR.size() != LR.size():
            scale = HR.size(2) // LR.size(2)
            LR = F.interpolate(LR, scale_factor=scale, mode="nearest")

Why don't you use bicubic?

Most people use bicubic in super resolution.
Do you have some special things?

I am interest in your CutBlur.
Thank you for your attention.

[nmhkahn]

First of all, in our internal analysis, we have found that the nn-based upsample is better than the bicubic (although it's marginal).
Unlike the early-upsample SR methods which mostly upsample via bicubic (e.g. VDSR, SRCNN), our intention of upsampling is not to reconstruct the image in high-res space, rather match the resolution to perform CutBlur.
In this respect, we have decided to provide the raw signal (pixel) itself by not using any high-level interpolation methods.

[LoveU3tHousand2]

So, the LR input , SR output and HR GT have same size ?
it's not for post-upsample SR methods? eg: LR.shape:120, SR.shape:480, gt.shape:480

[LoveU3tHousand2]

If LR and HR have same size at begin , after net.apply SR_size == 4 x LR.size, then how to calculate loss between SR and HR ? downscale HR or what ? Or just down(CutBlur(up(LR))) ?

[LoveU3tHousand2]

Umm, so the method is Cutblur(HR, LR) to replace HR rather LR ?

[LoveU3tHousand2]

... Forget it ... I just saw the first layer of these networks ...

[KaishanLam]

首先，在我们的内部分析中，我们发现基于 nn 的上采样优于双三次（尽管它是边际的）。 与主要通过双三次上采样的早期上采样 SR 方法（例如 VDSR、SRCNN）不同，我们上采样的目的不是在高分辨率空间中重建图像，而是匹配分辨率以执行 CutBlur。 在这方面，我们决定不使用任何高级插值方法来提供原始信号（像素）本身。

After cutblur, LR is now the same size as GT, do you want to downsample it by a factor of 4, f.object (LR, scale_factor=0.25, mode="nearest"), and bring it back to the small size?

[KaishanLam]

After cutblur, LR is now the same size as GT, do you want to downsample it by a factor of 4, F.interpolate(LR, scale_factor=0.25, mode="nearest"), and bring it back to the small size?