Towards Realistic Out-of-Distribution Detection: A Novel Evaluation Framework for Improving Generalization in OOD Detection

TL;DR

This paper introduces a new evaluation framework with realistic datasets and a generalizability score for OOD detection, revealing current models' limitations and proposing a post-processing method to improve performance under distribution shifts.

Contribution

It proposes new realistic OOD test datasets, a generalizability score, and a post-processing method to enhance pre-trained models' robustness in real-world OOD detection scenarios.

Findings

Existing benchmarks do not reflect real-world distribution shifts.

State-of-the-art pre-trained models perform poorly on new realistic datasets.

Post-processing significantly improves model performance under distribution shifts.

Abstract

This paper presents a novel evaluation framework for Out-of-Distribution (OOD) detection that aims to assess the performance of machine learning models in more realistic settings. We observed that the real-world requirements for testing OOD detection methods are not satisfied by the current testing protocols. They usually encourage methods to have a strong bias towards a low level of diversity in normal data. To address this limitation, we propose new OOD test datasets (CIFAR-10-R, CIFAR-100-R, and ImageNet-30-R) that can allow researchers to benchmark OOD detection performance under realistic distribution shifts. Additionally, we introduce a Generalizability Score (GS) to measure the generalization ability of a model during OOD detection. Our experiments demonstrate that improving the performance on existing benchmark datasets does not necessarily improve the usability of OOD detection models in real-world scenarios. While leveraging deep pre-trained features has been identified as a promising avenue for OOD detection research, our experiments show that state-of-the-art pre-trained models tested on our proposed datasets suffer a significant drop in performance. To address this issue, we propose a post-processing stage for adapting pre-trained features under these distribution shifts before calculating the OOD scores, which significantly enhances the performance of state-of-the-art pre-trained models on our benchmarks.