psnr/ssim
Closed this issue · 5 comments
Hi, which file do you use to compute the psnr/ssim of your results?
Thank you.
Kindly refer to #20 (comment)
Thank you for your reply. I notice that the code is for calculation of one attribute manipulation. The paper shows that the psnr/ssim is 31.67/0.948. What attribute or attributes are these two numbers point to?
By the way, I tested the attribute 'Eyeglasses' and used a test set of 6000 images from your test set. The result of your trained model is 26.524815741791713 / 0.9064622972518696. I also retrained on your training set following the paper's instruction and the retrained the model got a psnr/ssim of 27.0123/0.9045. I am wondering why these two gaps happens.
Thank you.
Hi Xijie,
The PSNR and SSIM are used to evaluate the reconstruction output images, since we have no groundtruth of manipulated images.
And the PSNR and SSIM on the reconstruction images are used to indirectly show the image quality.
What do you mean by reconstruction output images? The results of editing and editing again(reversion)?
No, the reconstruction output images are generated by att_diff=0 for STGAN, and att_target=att_source for others