用文档类屏幕照片跑模型生成全黑的图

[EYcab]

我之前已经调通你这边的模型并用来跑生成过程，我发现如果用 AIM2019 的测试集跑，生成的图片看起来挺理想的，但是当我用业务数据（文档类屏幕照片）来跑的时候就会生成全黑的图片了，这是个什么现象呢？

是不是模型不够鲁棒？

[zhenngbolun]

Perhaps the model weights trained on AIM2019 dataset is not robust enough to solve general demoireing problems because the AIM2019 dataset is a generated dataset (not a real dataset). Basing on our experience, the model weights trained on TIP2018 dataset will be more robust for real image demoireing.

However, as you say that the model outputs a black image seems impossible. Maybe you should check your code first. Please notice that the output of the model is image with the range of [0, 1]. I guess you didn't convert the image to [0, 255] and directly saved the image of [0, 1] to the disk.

[EYcab]

Indeed, I ran your codes via main_multiscale.py where already have 255.0 multiplied. That is why I don't think that messes it up.

[zhenngbolun]

We may upload the model weights trained on TIP2018 dataset later. You can try with this model weights.

[EYcab]

Thank you in advance,@zhenngbolun, but I have not seen your updated model weights on TIP2018, have you uploaded them yet?

[daeing]

IM2019 dataset is not robust enough to solve general demoireing problems because the AIM2019 dataset is a generated dataset (not a real dataset). Basing on our experience, the model weights trained on TIP2018 dataset will be more robust for real image demoireing.

I notice you said AIM2019 dataset is a generate dataset. Do you know how to generate this dataset?

[fxcdl]

HI, I test the model on AIM2019 also generated black image, and I find the model output the negative matrix，so after *255, it s turned all back.I think it s abnormal.

[AWangji]

我之前已经调通你这边的模型并用来跑生成过程，我发现如果用 AIM2019 的测试集跑，生成的图片看起来挺理想的，但是当我用业务数据（文档类屏幕照片）来跑的时候就会生成全黑的图片了，这是个什么现象呢？

是不是模型不够鲁棒？

同，您解决了吗