Semantic-Discrepant Outliers on CIFAR-10 Dataset

Suhee Yoon, Sanghyu YoonZenodo 2023 · DOI · OpenAlex

Abstract

We provide synthetic Out-of-distibution (OOD) dataset, which is called Semantic-Discrepant (SD) outliers, on CIFAR-10 dataset. SD outliers can be utilized for boosting OOD detection model performance. For the details, SD outliers are realistic OOD samples that contains incoherent semantic shift while preserving nuisances with in-distribution (ID). SD-outliers are generated from ID training samples using semantic-discrepant sampling in the diffusion model. so SD-outliers on CIFAR-10 contains 50000 32X32 images which is same as CIFAR-10 training dataset size. The dataset has a capacity of 768MB.
BibTeX
@inproceedings{yoon2023semantic,
  author = {Suhee Yoon and Sanghyu Yoon},
  title = {{Semantic-Discrepant Outliers on CIFAR-10 Dataset}},
  booktitle = {Zenodo},
  year = {2023},
  doi = {10.5281/zenodo.8394846},
}

Loading related content...

Figure