AiArt-Gao / HIDA

[ICCV'23] Human-Inspired Facial Sketch Synthesis with Dynamic Adaptation
MIT License
25 stars 4 forks source link

FID 和LPIPS #3

Open ludandandan opened 7 months ago

ludandandan commented 7 months ago

你好,使用pre-trained model 和dataset FS2K生成的结果(1037张素描图【由于test的photo只有1037张,这也是一个疑问点】)计算FID,分别使用了https://github.com/AiArt-HDU/HIDA/blob/main/evaluation/dirs_fid_score.py和https://github.com/mseitzer/pytorch-fid的方法计算的FID得分都是58左右(ground truth是dataset FS2K中的sketch),并且对FSGAN计算的FID(groud truth使用的是FS2K原版的图像,由于FS2K中图像尺寸不一,全部resize到256x256计算的)得分在44左右,使用https://github.com/richzhang/PerceptualSimilarity计算的LPIPS,HIDA方法得分是0.273(这里只有1037张图),而FSGAN计算的是0.228(这里是正常的1046张图),除了0.273这个数值外,其他数值与论文中的差距有些大,请问是我哪里处理不当吗?

zjyruobing commented 7 months ago

请问你们可以跑通训练的代码吗

ludandandan commented 7 months ago

我试了一下测试的代码,可以生成图像,训练的还没搞

JcccKing commented 7 months ago

The size of the image generated may be 250 * 250

Lun-menG commented 4 days ago

你好,请问你们的anno_train.json文件是在哪里获取的呢

ludandandan commented 4 days ago

https://github.com/DengPingFan/FS2K从这里下载的

Lun-menG commented 3 days ago

https://github.com/DengPingFan/FS2K从这里下载的 哦哦,好的,非常感谢