About train datasets

[Madeline-hyh]

Great work, I'm very interested in it. But I have a small question, I want to know what is the training dataset used for the wild-daclip_ViT-L-14.pt model, and do all the training datasets contain the four mixed degradations (blur, noise, resize, jpeg)? Is the wild degradation "blur, noise, resize, jpeg"? We look forward to your reply and thank you.

[Algolzw]

Thank you! We use the LSDIR dataset to train the model. For both daclip_ViT-L-14 and Wild-IR, all training LQ images are generated using the random_degradation function.

[Madeline-hyh]

Thank you very much for your patient reply. After reading the code file you gave, I found that your deg_list does not have resize (it is replaced by blur), so it is convenient to know why? Thank you.

[Algolzw]

Great! I prefer 'blur' since most 'resize' operations use bicubic/bilinear/nearest interpolation.

[Lincoln20030413]

插值会影响什么呢？这个操作不是之前的图像复原模型中造超分数据集时都会用的吗？期待您的解答

[Algolzw]

插值也可以看作是一种blur操作，在超分中一般假定是先blur再下采样，但这里我们跳过了下采样这一步（因为这个模型不支持超分任务）。

[Lincoln20030413]

感谢回答，那你们不用resize操作而用blur是因为resize本身的插值带来了一些问题吗？

[Algolzw]

我也不确定= =，不过有时候resize应该也是可以的，一般看任务而定。