Questions about the dataset

[X-funbean]

Hi, NICO++ is a wonderful dataset for DG and OOD researches, but I'm confused at the training and testing protocols of it.

In your paper NICO++: Towards Better Benchmarking for Domain Generalization, the track1, i.e. classical DG problem is evaluated by training on 6 domains and then testing on the other 4 domains, as showed in table 5:
.

However, the downloaded dataset only contains 6 domains, as showed in the figures below:

I wonder if this is the full version of NICO++. Am I missing anything to reimplement the results in the paper.

On the other hand, I notice that there is a 'public_test_flat' folder along with the 'train' folder, with 13907 images without meaningful file names. Are these images used for testing? If so, how to evaulate the results without groundtruth labeling?

Looking forward to your reply.

[BrianXiao1]

Same Question. Where are the rest 4 domains? lol
@X-funbean 哥 这个问题你解决了吗