Did you use the exact same train/test split as Pose-Transfer?

[Ha0Tang]

The results (IS, SSIM, and DS) of Table 1 in your paper is quite different from the results of Table 1 in Pose-Transfer. For example, you reported 2.265 (IS) and 0.770 (SSIM) on Def-GAN, while Pose-Transfer reported 3.362 (IS) and 0.760 (SSIM) on Def-GAN. I would like to know what makes these differences? Thanks in advance!

[menyifang]

Hi @Ha0Tang, we follow the train/test split in PATN, which is different from Def-GAN, making these differences. However, we found there is a typo for the IS of Def-GAN(should be 3.265, not 2.265) after checking. Sorry for this mistake and we will correct it in our arxiv version.

[Ha0Tang]

Hi, thank you for your quick response. You reported 3.265 (IS) and 0.770 (SSIM) on Def-GAN, while Pose-Transfer reported 3.362 (IS) and 0.760 (SSIM) on Def-GAN. Since both of you used the same testing set on Def-GAN, I would like to know what makes these differences? Also, you reported 3.209 (IS) and 0.774 (SSIM) on Pose-Transfer, while in their paper the authors reported 3.209 (IS) and 0.773 (SSIM), I would like to know why IS is the same and SSIM is slightly different? Thanks.

[menyifang]

@Ha0Tang, I think it may come from that the reproduction details for other papers (e.g. Def-GAN) between us and Pose-transfer are not entirely the same. Also different pytorch or tensorflow versions could be used.