A Benchmark and Evaluation for Real-World Out-of-Distribution Detection Using Vision-Language Models

Shiho Noda, Atsuyuki Miyai, Qing Yu, Go Irie, Kiyoharu Aizawa

Out-of-distribution (OOD) detection is a critical safety feature for machine learning models, yet existing benchmarks often fail to capture the complexity of real-world scenarios. This paper introduces three new hierarchical benchmarks—ImageNet-X, ImageNet-FS-X, and Wilds-FS-X—designed to evaluate Vision-Language Models (VLMs) against fine-grained semantic and covariate shifts. Through extensive evaluation of CLIP-based methods, the authors demonstrate that high classification accuracy does not guarantee robust OOD detection. Notably, few-shot tuning can introduce covariate biases that degrade performance in changing environments. These findings highlight the need for OOD detection methods that can decouple semantic understanding from stylistic variations.

The Challenge of Real-World OOD Detection

In machine learning, "knowing what you don't know" is essential. A model trained on dogs should identify a car as "unknown" rather than misclassifying it. However, current OOD benchmarks have largely reached performance saturation because they rely on visually distinct categories. In reality, models often struggle with "hard" cases: distinguishing between similar species (semantic shift) or handling images with different lighting and backgrounds (covariate shift). This paper addresses the lack of challenging, real-world benchmarks that test these subtle but vital distinctions.

Examples of 'hard' semantic shifts in the ImageNet-X benchmark, where ID and OOD classes are closely related.

Hierarchical Benchmarks for VLMs

To address these gaps, the researchers developed three benchmarks using the WordNet hierarchy to ensure semantic proximity between In-Distribution (ID) and Out-of-Distribution (OOD) classes:

  1. ImageNet-X: Focuses on "hard" semantic shifts by splitting the 1,000 ImageNet classes into 500 ID and 500 OOD classes that are closely related (e.g., "dalmatian" is ID, while "Great Pyrenees" is OOD).
  2. ImageNet-FS-X: A "full-spectrum" benchmark that adds covariate shifts—such as corruptions and stylistic variations—to the semantic splits of ImageNet-X using variants like ImageNet-R and ImageNet-C.
  3. Wilds-FS-X: Features real-world data from the WILDS collection, including wildlife camera traps (iWildCam) and satellite imagery (FMoW), which naturally contain shifts in location, time, and equipment.
Overview of the hierarchical benchmarks: ImageNet-FS-X and Wilds-FS-X (iWildCam and FMoW), illustrating both semantic and covariate shifts.

Key Insights and Results

The study evaluated various CLIP-based strategies, including zero-shot methods and few-shot prompt tuning. Several key findings emerged:

  • The Accuracy Paradox: Improved classification accuracy does not necessarily lead to better OOD detection. In the iWildCam tests, few-shot learning significantly boosted classification for known animals but actually reduced the model's ability to identify unknown species compared to zero-shot versions.
  • Covariate Bias: Few-shot tuning tends to specialize models to the specific lighting or camera style of the training set. This "covariate bias" causes OOD detection performance to plummet when the image style changes, even if the semantic content remains consistent.
  • Performance Drop: On traditional benchmarks, models often score 90-95% AUROC. On the new ImageNet-X, scores fall to 70-80%. The FMoW satellite benchmark proved most difficult, with scores near 52-56%—barely exceeding random chance.

Conclusion

The results suggest that current OOD detection strategies are not yet robust enough for real-world deployment. The "saturation" seen in older benchmarks is misleading, as it fails to account for the subtle semantic and covariate shifts encountered in the wild. Future research must focus on developing methods that can decouple semantic substance from covariate "style" to ensure models remain reliable across diverse environments.

Summary updated

Abstract

Out-of-distribution (OOD) detection is a task that detects OOD samples during inference to ensure the safety of deployed models. However, conventional benchmarks have reached performance saturation, making it difficult to compare recent OOD detection methods. To address this challenge, we introduce three novel OOD detection benchmarks that enable a deeper understanding of method characteristics and reflect real-world conditions. First, we present ImageNet-X, designed to evaluate performance under challenging semantic shifts. Second, we propose ImageNet-FS-X for full-spectrum OOD detection, assessing robustness to covariate shifts (feature distribution shifts). Finally, we propose Wilds-FS-X, which extends these evaluations to real-world datasets, offering a more comprehensive testbed. Our experiments reveal that recent CLIP-based OOD detection methods struggle to varying degrees across the three proposed benchmarks, and none of them consistently outperforms the others. We hope the community goes beyond specific benchmarks and includes more challenging conditions reflecting real-world scenarios. The code is https://github.com/hoshi23/OOD-X-Benchmarks.
BibTeX
@misc{noda2025benchmark,
  author = {Shiho Noda and Atsuyuki Miyai and Qing Yu and Go Irie and Kiyoharu Aizawa},
  title = {{A Benchmark and Evaluation for Real-World Out-of-Distribution Detection Using Vision-Language Models}},
  year = {2025},
  eprint = {2501.18463},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.18463},
}
Figure