search

Lotayou / Face-Renovation

Official repository of the paper "HiFaceGAN: Face Renovation via Collaborative Suppression and Replenishment".
https://arxiv.org/abs/2005.05005
GNU General Public License v3.0
284 stars 49 forks source link

niqe #14

Closed dan520520520 closed 4 years ago

dan520520520 commented 4 years ago

你好，可以透漏一下您输入niqe计算相似度的人脸图像尺寸以及 patch size大小吗？

Lotayou commented 4 years ago

Here https://github.com/Lotayou/Face-Renovation/blob/7362555cf681981e0906d9bd89f3ded4fdfbd75f/metrics_package/niqe/niqe.py#L197-L201

dan520520520 commented 4 years ago

谢谢，如果输入是68*68，您建议的patch size是？

dan520520520 commented 4 years ago

谢谢，如果输入是6868，您建议的patch size是？ 这两天测试了把人脸裁剪到6868，patch size = 15，好像效果还不错， 但是112*112好像没有找到合适的参数，关于NIQE使用到人脸图像上面，您有什么建议吗？

------------------ 原始邮件 ------------------ 发件人: "Lotayou/Face-Renovation" <notifications@github.com>; 发送时间: 2020年8月7日(星期五) 下午4:28 收件人: "Lotayou/Face-Renovation"<Face-Renovation@noreply.github.com>; 抄送: "丹丹"<1049011336@qq.com>;"Author"<author@noreply.github.com>; 主题: Re: [Lotayou/Face-Renovation] niqe (#14)

Closed #14.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

Lotayou commented 4 years ago

不建议输入裁到68*68，在这么低的分辨率下人脸细节基本全损失了，用NIQE不太好评价细节的生成质量