search

BadourAlBahar / pose-with-style

[SIGGRAPH Asia 2021] Pose with Style: Detail-Preserving Pose-Guided Image Synthesis with Conditional StyleGAN
https://pose-with-style.github.io/
MIT License
268 stars 41 forks source link

Different PSNR and SSIM #15

Closed samaonline closed 1 year ago

samaonline commented 2 years ago

Hello, thanks for the great work! I'm trying to reproduce the PSNR and SSIM with Global-Flow-Local-Attention. With the pre-trained model, it's 17.7 and 0.70; with the images you provided here, it's 17.7 and 0.69. They are worse than reported in the paper.

I wonder if the difference is negligible or if the metric used is problematic? Thanks!

metrics

BadourAlBahar commented 2 years ago

I am not really sure. However, please note that we only compute the PSNR/SSIM of the foreground human. We do so using the silhouettes provided here.

samaonline commented 1 year ago

Thanks! After using the masks, the metrics roughly match.