Question on splits of dataset DOSCO

[hxynjk]

Hi! Thanks for this great contribution to domain generalization. I note that you supply the domain labels of each sample in DOSCO. Besides, you also supply the random split of "train", "valid", "test" set. I want to know if you consider the domain labels while splitting the "train", "valid", "test" set.

Meanwhile, I note that the domain labels vary from 0 to 9, which makes it hard for us to split samples of each domain separately. If we do this, there will be so few samples in some domains that we can not train the network with enough samples.

Thanks for your reply! Thanks for sharing the public dataset.

[KaiyangZhou]

The domain labels in DOSCO are not supposed to be used. You can treat DOSCO like "ImageNet -> ImageNet-V2 / Sketch / Rendition"