Dataset Construction in BID Model's Task II.A

[Adele0108]

Thank you for your excellent work on the BID model and for making it publicly available. I have a few questions about the dataset construction in Task II.A:

(1) The clean CityScape training set contains 2975 images, but the provided masks have different counts (rain streaks: 1620; raindrops: 3500; snow: 3500). Could you clarify how these masks are applied given the count mismatch?

(2) Some masks don't match CityScape's standard 1024×512 resolution. How are these different-sized masks properly aligned with the base images?

(3) Would it be possible to release the complete multi-weather dataset to facilitate reproducibility?

(4) Currently, raina_dataset.py doesn't support testing individual weather cases separately. Could you please provide the dataset code for testing individual weather cases separately?

I greatly appreciate your time and consideration of these questions. Looking forward to your reply!

[JunlinHan]

Hi, thanks for your nice words about our work.

(1) The clean CityScape training set contains 2975 images, but the provided masks have different counts (rain streaks: 1620; raindrops: 3500; snow: 3500). Could you clarify how these masks are applied given the count mismatch?

A: We simply ignore the difference as datasets are generated on the fly. So there might be repeated use.

(2) Some masks don't match CityScape's standard 1024×512 resolution. How are these different-sized masks properly aligned with the base images?

A: Simply resize to the same resolution.

(3) Would it be possible to release the complete multi-weather dataset to facilitate reproducibility?

A: We generate the training set on the fly (to increase diversity), you can find the whole generation codes in this repo.

(4) Currently, raina_dataset.py doesn't support testing individual weather cases separately. Could you please provide the dataset code for testing individual weather cases separately?

A: Testing and training with single case can be supported. See: https://github.com/JunlinHan/BID/blob/59b7037e64a6b5ad2348d26242d04e0cb6951a9d/models/raina_model.py#L23C51-L24C1