About Table 2 in the paper, I wonder how much effect does the clean module and clean loss have towards the result?

[CrissyHoo]

The author gives the perceptual effect with different clean module and loss in figure 4. But these are not the quantitative results correspondingly. I mean, in Table 2, BasicVSR++ item shows the results training with bicubic. How about the results of BasicVSR++ training with the degradation data produced by "second-order order degradation model"? And why BasicVSR++ is used as the reference but not the BasicVSR?

[ckkelvinchan]

We did not try BasicVSR++ with second-order degradation. We use BasicVSR++ because it is the best model on bicubic degradation. It is just to provide a reference point to show how the best bicubic model works on real-world videos.

[ckkelvinchan]

I forgot that we did a very rough experiment on BasicVSR++

We trained BasicVSR++ for a very short period of time using the second-order degradation scheme. It is indeed better than the one trained with bicubic downsmapling. It works on some examples, but similar artifacts are observed.

Please note that this is not a complete experiment, and we should not draw any conclusion from the above experiment.

[Magauiya]

Hello Mr. Chan,

• Can you please share quant. results (Table 2) for the BasicVSR trained on exactly the same settings as RealBasicVSR (dataset degradation model, number of iterations)?
• Are those methods mentioned in the Section 5.1 (e.g. DBVSR and RealVSR) trained on the same dataset with degradations as the proposed one?

Thanks!

[ckkelvinchan]

Hello Mr. Chan,

• Can you please share quant. results (Table 2) for the BasicVSR trained on exactly the same settings as RealBasicVSR (dataset degradation model, number of iterations)?
• Are those methods mentioned in the Section 5.1 (e.g. DBVSR and RealVSR) trained on the same dataset with degradations as the proposed one?

Thanks!

1. We did not train BasicVSR with GAN. So I am afraid I do not have this result.
2. For DBVSR, it is trained on REDS, but the degradations should be different. For RealVSR, it is trained on their proposed dataset. The model they use is EDVR.