search

aim-uofa / Framer

Official PyTorch implementation of "Framer: Interactive Frame Interpolation".
206 stars 7 forks source link

Different result in VIDM paper #4

Closed Yolice closed 1 day ago

Yolice commented 2 days ago

I note that in the VIDM paper they report PSNR and SSIM for mid-frame scores, and the LPIPS has significant differences from your results. May I ask if your experiment is testing the Davis-7 dataset being cropped at 256*256 size?

encounter1997 commented 2 days ago

Hi, thanks for your interest!

The DAVIS-7 evaluation dataset is cropped at 256x256 resolution, using the official code here. Besides, we use the LPIPS implementation here.

Yolice commented 2 days ago

Hi, thanks for your interest!

The DAVIS-7 evaluation dataset is cropped at 256x256 resolution, using the official code here. Besides, we use the LPIPS implementation here.

hi,thanks for your reply. Can I ask which network you used, VGG or Alexnet, and what packages the SSIM and PSNR metrics were calculated from (e.g. Skimage?).

encounter1997 commented 2 days ago

Certainly, we use VGG for LPIPS calculation. We follow the PSNR and SSIM implementation in AMT repo here.

Yolice commented 2 days ago

Certainly, we use VGG for LPIPS calculation. We follow the PSNR and SSIM implementation in AMT repo here.

Thanks for your reply, May I ask again if the FID and FVD calculation scripts are available. Thank you very much

encounter1997 commented 2 days ago

We use pytorch-fid for FID calculation and followed TATS for FVD calculation.