zzh-tech / animation-from-blur Goto Github PK

I am trying to use your approach to remove the motion blur of a dance video, like in your examples, with a single person looking to the camera all the time, no background blur or movement, but im not sure how to do it.

I have placed the checkpoints, I have changed the dataset config to point to a single folder from my video, that i called it test, and i have placed the frames of my video with the correct names on "blur" folder. The rest of folders from the dataset I dont know how to generate them.

Can you give me some directions of how to do it please?

Basicly what i have done

• Editted the config to point just my test folder
• Place in dataset my video dataset, like
  dataset/test/blur -> my video frames
  dataset/test/sharp -> empty
  dataset/test/blur_anno -> empty
  dataset/test/trend+_avg -> empty
• Place your checkpoints

What i want its giving my video, use the checkpoints to correct the blur of my video

Thank you in advance.

Hellow !!
Thanks for your great work and implementation!

Can you please provide the result images on GOPRO testset or test codes?

When experimenting with the pretrained model you provide (Multi-modal blur decomposition evaluation with guidance predictor), the result images on GOPRO dataset are differ from those in the paper.
However, I don't know why the result images are different from those reported in the paper.

• I prepared each blurry image by averaging 7 consecutive sharp frames from the GOPRO test set.

• Example of input image
  

• Result image I get (this is the first frame result image)
  

Hello!!
Thanks for your great work and implementation!

When I was downloading B-AIST++ Dataset , there are always problems.
When I was halfway through the download, the browser showed a network error and the download failed.
I tried many times, but the download failed.
Do you have any other download connections?

Hello @zzh-tech ,
Thanks for sharing the codes. I want to retrain the model on the GoPro/GenBlur dataset. It seems that there should be three elements including blur, sharp, and trend+ in the dataset. But I don't know how the dataset is generated from the original GoPro dataset (especially the trend+ element). Could you please share the code for dataset preparation or provide the generated dataset GenBlur mentioned in the paper? Thanks a lot!

    Dataset class of videos, each video must be represented as a dir
    Structure under each video dir:
        blur: {:08d}.png (start from 00000000.png)
        sharp: {:08d}_{:03d}.png (start from 00000000_000.png)
        trend+: {:08d}_trend.npy (start from 00000000_trend.npy)