Question about dataset preparation

[tongzixin716716]

Hi, thank you for releasing this code.

I am trying to follow the dataset format used in this repository, where each dataset is organized as:

<dataset>/<train or test>/<domain>/<class>/

I would like to ask whether you could provide more details about the dataset preparation process, especially for datasets whose official releases do not naturally follow this structure.

In particular, I am confused about the preprocessing for:

• OfficeHome
  The official dataset seems to be organized by domain/class, and in many DG settings, the protocol is usually to treat some domains as source/train domains and one or more domains as target/test domains, rather than splitting each individual domain into its own train/test subset. Could you please clarify how this train/test split was created for each domain in OfficeHome?

• NICO++
  The official release I downloaded contains train and public_test_flat, but public_test_flatpart seems to contain images without class-organized labels, so it does not directly match the repository format. Could you please explain how the NICO++ test set was prepared in your codebase?

• TerraIncognita
  The official annotations are provided as global splits (e.g. train, cis, trans), so if I directly convert these official global splits into the repository format, I encounter a problem where some classes (or even some domains) end up with no samples in the train split. Could you please clarify how TerraIncognita was processed?

If there are any preprocessing scripts, split files, or additional instructions for these datasets, could you please share them?

That would be very helpful for reproducing the experiments. Thanks a lot.

[VirtualSpaceman]

Hi, sorry for the late reply. I didn't receive any notifications about this issue, so I apologize for the delay.

I've uploaded the preprocessing scripts for each dataset under scripts/datasets/. At the top of each file, you'll find the download link, along with details on how the train/test splits are created.

• For OfficeHome, each domain is split independently with an 80/20 random split. The resulting files are then copied into train/<domain>/<class_name> and test/<domain>/<class_name> folders. So the train/test split is created separately for each domain.
• NICO++ uses the images under public_dg_0416/train/<domain> as input. We consider all images within this folder then applies the 80/20 random split per domain. Finally, we organize the images of each split into folders that follows train/domain/<class_name> and test/domain/<class_name> format.
• In TerraIncognita, we merge the train+val splits and use it as train and keep the keep the test untouched. Again, the files are then organized to follow the folder structure train/<location_x>/<class_name>] and test/<location_x>/<class_name>