search

ckkelvinchan / BasicVSR_PlusPlus

Official repository of "BasicVSR++: Improving Video Super-Resolution with Enhanced Propagation and Alignment"
Apache License 2.0
568 stars 63 forks source link

denoise pretrained model didn't work #12

Closed bjiale closed 2 years ago

bjiale commented 2 years ago

Hi @ckkelvinchan , I have tried to run the denoise pretrianed model with DAVIS-test dataset, but the result images has no difference between the input images. My experiment is as below: pretrained model: basicvsr_plusplus_denoise-28f6920c.pth optical flow model: I download it to my local path dataset: I use the the tractor folder in DAIVS-test dataset which is used in your paper. I add the sigma 50 gaussian noise to the pictures as input. result: image

It seem that it run successfully, but the result has no change between the input. Is their any parameter or option error in my attempt? By the way, basicvsr_plusplus_c64n7_8x1_600k_reds4.pth can run successfully and the result is excellent!

ckkelvinchan commented 2 years ago

That’s strange. Can you send me the input you used and let me have a try?

bjiale commented 2 years ago

My input data. tractor.zip thanks~

ckkelvinchan commented 2 years ago

How did you add the noise? Also, I observed pixel shift when compared with the GT.

bjiale commented 2 years ago

I used the RandomNoise function in your code to add the noise. image

bjiale commented 2 years ago

About the pixel shift, I crop 2 pixels in width to make the image size can divided by 4.

bjiale commented 2 years ago

Is my data's noise type different from the training dataset?

ckkelvinchan commented 2 years ago

I guess that's because you saved the frames as .jpg, which produces compression. Try saving it to .png.

In addition, For frames not divisible by 4, I padded zero at the end (i.e., last rows and columns) and crop it after processing, to make sure there is no pixel shift.

bjiale commented 2 years ago

Yes, thank you~ You are right. I try to use png format image and it works!