EDRN can not train successfully

[HymEric]

When I try to train EDRN, encountered this: it seems the dimension is not match.

[LoSealL]

EDRN is an image SR network. Seems you are trying to train with video data?

[HymEric]

I really use the image data not video data. I just change the default scale=1 to scale=4 before train.

[LoSealL]

Oh, I see. EDRN is from NTIRE 2019 challenge - Real SR, where its purpose is filtering an image from blur to sharp. EDRN itself doesn't contain any upsample module, so changing scale from 1 to 4 needs to modify inside the architecture.

If you are searching for an SOTA image SR network, you can use ESRGAN and set weights to [1, 0, 0].

[HymEric]

But I saw the NTIRE 2019 Challenge on Real Image Super-Resolution: Methods and Results EDRN is a SR model. In your NTIRE it seems like including the SR method and denoising method but I couldn't use it correctly.

[LoSealL]

NTIRE 2019 SR is a 1:1 scaling challenge

[HymEric]

Do you say this competition.

[LoSealL]

Of course. I participated in the challenge and got 13th place in the final.

[HymEric]

I'm still confused. The competition conclusion paper said EDRN proposed by IVIP-LAB team got 9th place. And the architectures in paper have upsample operator to get HR.

[LoSealL]

You misunderstand, we all use U-net architecture, where in the encoder (at head), pooling to a downsampled feature maps and in the decoder (at tail) upsampled to normal size. The overall resolution is not changed.

[HymEric]

Yes, you are right. But can you tell me why it say result output as HR many times in the conclusion paper ? I always think HR is the output of SR work.

[LoSealL]

Because in the challenge, we just call "Ground-Truth" as "HR", and its counterpart as "LR". LR looks more blurry and HR looks clear and sharp. Maybe a little confused :(

[HymEric]

Oh. Thank you very much!