GENERALIZED OUTLIER EXPOSURE FOR ROBUST OUT-OF-DISTRIBUTION DETECTION: A STUDY ON MIXING, FILTERING, AND AUGMENTATION

Date

2025-8

Author

Öztel, Fatma

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

4550
views

0
downloads

Cite This

Outlier Exposure (OE)-based Out-of-Distribution (OOD) detection methods achieve strong performance by leveraging OOD data during training. However, OE can sometimes negatively impact classification performance, which remains an ongoing challenge. This thesis evaluates Generalized Outlier Exposure (G-OE) methods, which introduce two strategies—smoothing the confidence surface with Mixup and filtering confusing OOD samples—to mitigate the negative effects of OE-based OOD detection. In addition to assessing the performance of Mixup and filtering, various alternative Mixup, filtering, and augmentation techniques are investigated to develop more robust OOD detection methods. CutMix and Manifold Mixup are explored as alternatives to Mixup; uncertainty-based (margin, entropy) and feature-based (cosine, Euclidean) filtering methods are examined as alternatives to conventional filtering; and diverse augmentations—including geometric and photometric transformations—are implemented. The effects of these methods on both classification accuracy and OOD detection performance are evaluated through extensive experiments. As a result, the strengths and limitations of these approaches are discussed, providing valuable insights for the literature.

Subject Keywords

Out-of-distribution Detection, Outlier Exposure Based OOD Detection, Mixup, Filtering, Augmentation

URI

https://hdl.handle.net/11511/115588

Collections

Graduate School of Informatics, Thesis