Questions about the degradation process for BDx4 low-resolution input

[yhjoker]

For BIx4, we use MATLAB to resize the images: lq = imresize(gt, 0.25).

For BDx4, you can find the MATLAB script here.

Originally posted by @ckkelvinchan in #487 (comment)

[yhjoker]

Hello there, I am currently trying to train the basicvsr model on Vimeo-90K( BD degradation ) and find that there are slight differences between the downsampling functions used in DUF( BD function in EDVR repo can be found here ) and those used in Basicvsr ( for example, gussian smooth kernel size ). Is there any reason why we choose to use the Matlab version in basicvsr training?

[ckkelvinchan]

The sigma we used is the same. About the kernel size, I followed previous works (I am not sure whether I copied from RCAN but RCAN uses this setting here).

[yhjoker]

The sigma we used is the same. About the kernel size, I followed previous works (I am not sure whether I copied from RCAN but RCAN uses this setting here).

Thanks for the answer！ @ckkelvinchan I am trying to retrain the basicvsr model using the degradation function provided by EDVR ( here is the link ) but find the training process can not convergence. I use the pretrained network-sintel-final.pytorch model for SPynet at here and basicvsr_vimeo90k_bd.py config file( only spynet_pretrained path is modified ), is there any tools to help locate the training problems in mmediting repo?

[ckkelvinchan]

Could you provide me with more information? For example, the loss curve and how you generate the data.

[yhjoker]

Could you provide me with more information? For example, the loss curve and how you generate the data.

After adjusting the learning rate to 1e-4, things start to get better. I use the degradation function proposed by DUF to generate LR input( the duf_downsample provided in EDVR repo ). Other parameters might influence the training process are:

• learning rate : 1e-4
• samples_per_gpu : 2
• gpus : 4
• total_iters : 300000

It seems loss is decreasing and validation psnr is increasing( very unstable, though ). The loss curve and validation results for 250000/300000 training process is showed as following. Maybe a larger total_iter setting is required to get a promising result. Thanks again for your kind reply!

[ckkelvinchan]

I think a PSNR of about 25 is a bit low. Did you compare this result to the output upsampled by bicubic interpolation?

[ckkelvinchan]

Closing now due to inactivity. Please reopen if the problem persists.