about Rain100L dataset

[jimmy820904]

Hi. Does Rain100L training dataset only has 200 images?
Rain100L in other paper seems to have 1800 images for training.
Does this problem affect the experimental results?

[csdwren]

The authors of Rain100H and Rain100L have updated the datasets. This paper is based on the original datasets.
For the results on new datasets of Rain100H and Rain100L, please refer to https://github.com/csdwren/RecDerain (RainHeavy* and RainLight*)

[csdwren]

[jimmy820904]

The authors of Rain100H and Rain100L have updated the datasets. This paper is based on the original datasets.
For the results on new datasets of Rain100H and Rain100L, please refer to https://github.com/csdwren/RecDerain (RainHeavy* and RainLight*)

Thanks for your reply!
I have another question.
If I want to get a good deraining result, I should put the three datasets together for trainging, right?

[csdwren]

(1) For the individual training dataset, the model fully trained for Rain100L or RainLight* has the best generalization ability to real rainy images.

(2) As for the PReNet for real images, I put the three datasets together. But Rain1400 dataset has much more images than the other two. So when preprocessing, the stride for extracting training patches from Rain1400 is larger. Overall Rain100H plays the more important role, but it is quite different from real rain streaks. I found PReNet trained after 1 epoch has better generalization ability for real images than the model trained after 100 epoches. That is, PReNet after 100 epoches is overfitted to training dataset. PReNet after 1 epoch can better remove rain streaks, but the deraining images are likely to be over-smoothed with less textures. I do not compare the model with (1). Maybe (1) is better.

[jimmy820904]

Thanks!