S-DCVROOD

Software, experiments and results of the paper DCV-ROOD Evaluation Framework: Dual Cross-Validation for Robust Out-of-Distribution Detection

Usage

Download datasets:

• download_datasets.py Some version cannot download emnist with torch, in that case download in a different way.

Create folds:

• create_folds.py
• create_folds.superclasses.py

Create ground truth folds:

• create_gt_folds --seed $1 --number $2
  • Seed indicate a seed for reproducibility
  • number is a seed to create diferent folds

Train:

• trainer.py This code has multiple parameters to train with diverse datasets, also it is posible to select a previous trained net and further training with outlier exposure.

Embbedind extractors:

• process_embs.py Extract the embeddings and logits of a trained net.

OOD detectors:

• exec_pp.py

• gt_exec_pp.py Given a trained net, an OOD detector, an ID dataset and an OOD dataset, these files calculate an OOD score.

• exec_pp_embs.py Instead of a trained net you can use the embbedings and logits to calculate an OOD score (if possible).

Reading:

• read_confidence.py
• convergen.py Calculate metrics given an OOD score.

Full experiments and results

If you are interested in reviewing the complete results of the experiments conducted in the paper, you can access them through this link.

Citation

If you find this work useful in your research, please consider citing:

@article{urrea2025dcv,
  title={DCV-ROOD Evaluation Framework: Dual Cross-Validation for Robust Out-of-Distribution Detection},
  author={Urrea-Casta{\~n}o, Arantxa and Segura-Kunsagi, Nicol{\'a}s and Su{\'a}rez-D{\'i}az, Juan Luis and Montes, Rosana and Herrera, Francisco},
  journal={arXiv preprint arXiv:2509.05778},
  year={2025}
}

(Pre-print citation, to be updated upon publication.)