search

deeplearning-wisc / MOOD

Code for CVPR2021 paper: MOOD: Multi-level Out-of-distribution Detection
38 stars 5 forks source link

Questions on OOD Datasets #2

Open superctj opened 3 years ago

superctj commented 3 years ago

Thank you for open-sourcing your code! I have a few questions about OOD datasets.

  1. As far as I know, PyTorch provides the SVHN dataset. Is there any difference between the one you use and the PyTorch version?
  2. I noticed that STL-10 is also in the collection of OOD datasets. But it seems there is a significant overlapping between classes of CIFAR-10 and STL-10. Could you share any insights into why STL-10 could be considered an OOD dataset?
myhakureimu commented 3 years ago

Thank you very much for flagging the issue! (1) We followed the robust out-of-distribution detection for SVHN. We’ve also tried SVHN from the pytorch version. They both works, both can be tried if curious. (2) We agree STL-10 may contain overlapping classes and will remove that. Thank you for your comment again!