Hi all,
I am confused about the training and validation dataset for the OOD task:
In the paper, it is mentioned that classes train ( label Id = 16), motorcycle ( label Id = 17), and bicycle (label Id = 18) have been removed to make the OOD dataset.
However, in the training and validation set I see images having label id:18. And, in all the images, label ID: 12,15 are missing.
Could you please help me understand the dataset? Maybe I could be missing something.
-Shyam
Posted by: Shyam671 @ Aug. 7, 2023, 1:52 p.m.Dear Shyam,
Thank you for leaving the message.
We are aware that there are a few cases with OOD label id:18 that exist in the training and validation set. Yet the amount of them is very small throughout these two sets (<500 pixels in total), and this also happens on real datasets.
Concerned about the examples with label id:12 (rider) and 15 (bus), we have bicycles and motorcycles as OOD examples, so we don't have riders on the training and validation sets. Meanwhile, because the semantic information of trains or trams is very close to that of buses, we don't have examples for class bus in the dataset, which is also mentioned in the paper of the dataset.