[Feature] Add FungiTastic as a fine-grained near-OOD benchmark

[michalneoral]

Hi @Jingkang50, I represent the original authors of FungiTastic and I'd like to propose
adding it to OpenOOD as a new ID + near-OOD benchmark. CONTRIBUTING.md does not describe a
procedure for contributing a new dataset, so I'm opening this issue to discuss it first.

I have access to the full dataset (it's hosted on our server), so if there's interest I'm
happy to contribute the integration as a PR and provide the data, splits, and a reference
checkpoint in whatever form you need.

What it is

FungiTastic is a published fine-grained fungi classification dataset and benchmark (~350k
observations, ~5k species; arXiv:2408.13632). It ships an explicit open-set task: a held-out
set of novel species that never appear in training. That gives a
real, in-domain near-OOD rather than a synthetic class split.

Why it is a strong fit for OpenOOD

• Natural near-OOD by construction: unseen species, same domain.
• Large and modern: ~2829 known classes, ~92k known + ~5.7k novel test samples.
• Adds a biology / fine-grained axis to OpenOOD, similar to how OpenMIBOOD (CVPR'25) added
  a medical one.
• The dataset is already published, so no new dataset paper is needed, just the integration.

Most of the integration is already prepared

While running OpenOOD's own postprocessors on FungiTastic, I put most of the integration
into the OpenOOD format already:

• Imglists for all splits.
• configs/datasets/fungitastic/*.yml and the DATA_INFO entry, mirroring imagenet200.

A PR would therefore mostly come down to agreeing on the layout and hosting.

Questions for you

1. Would you be interested in including FungiTastic in OpenOOD?
2. Data hosting: the dataset is on our server and I can serve an OpenOOD-formatted bundle
   (imglists + images) wired into scripts/download/download.py, or link the official
   FungiTastic download. Please let me know if you would prefer another arrangement.
3. Do you need a reference ID classifier checkpoint for this benchmark? If so, I can train
   and share one in whatever recipe you prefer.
4. Should this sit under the OSR / near-OOD benchmarks, or as a standalone fine-grained entry?

I believe FungiTastic fills a gap in the current benchmark suite: a large, real-world,
fine-grained near-OOD task where the OOD samples are genuinely novel classes from the same
domain, which is exactly the setting that matters in practice and is hard to obtain from
synthetic splits. I'd be glad to discuss the layout and open the PR whenever it suits you.

Best regards,
Michal