search

cuiaiyu / dressing-in-order

(ICCV'21) Official code of "Dressing in Order: Recurrent Person Image Generation for Pose Transfer, Virtual Try-on and Outfit Editing" by Aiyu Cui, Daniel McKee and Svetlana Lazebnik
https://cuiaiyu.github.io/dressing-in-order
Other
507 stars 126 forks source link

FID value less than paper #47

Closed rishabhjain9619 closed 2 years ago

rishabhjain9619 commented 2 years ago

Hi @cuiaiyu ,

Thanks for sharing your work. I was trying to reproduce the FID values for 256*256 images given in the paper using your code. However, I got a lesser FID value of 12.25 instead of 13.10 mentioned in paper.

For glfa also, I am getting a value of 9.87 instead of 10.57 which you also mentioned in gfla comment. Can you let me know how was that issue solved?

cuiaiyu commented 2 years ago

It seems a training set downsampling/compression problem. when computing FID statistics, GFLA seemingly uses a set of special parameters to preprocess the training/validation set. To reproduce GFLA result and to enable a fair comparison, GFLA's author kindly shared their training set and validation set (with preprocessed downsampling) with us, and that's how we get our numbers for 256x256.

We also get the better set of results (12.25 for us and 9.87 for GFLA) as yours by directly loading the highres images and downsampling by bilinear to compute the FID statistics.

rishabhjain9619 commented 2 years ago

Thanks for the quick reply. Can you share the code for downsampling the images or the post-processed images?

cuiaiyu commented 2 years ago

I am not sure if we are allowed to share the data that GFLA's authors shared with us. You may want to reach out to GFLA's authors for this. :)

rishabhjain9619 commented 2 years ago

Sure, no issues, thanks for the clarifications