C-WOE: Clustering for Out-of-Distribution Detection Learning With Wild Outlier Exposure.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Out-of-distribution (OOD) detection is vital for anomaly handling in computer vision.
• Outlier Exposure (OE) is effective but requires clean auxiliary OOD data, which is often infeasible.
• Wild outliers, abundant and easily obtained, offer potential but contain mixed in-distribution (ID) and OOD samples, posing supervision challenges.

Purpose of the Study:

• To develop an effective strategy for leveraging wild outliers in OOD detection.
• To mitigate the negative impact of in-distribution samples within wild outlier datasets.
• To improve the reliability and performance of OOD detection systems.

Main Methods:

• Proposed Clustering for Wild Outlier Exposure (C-WOE) method.
• Dynamically reweighting samples within wild outliers to assign higher weights to OOD samples and lower weights to ID samples.
• Theoretical guarantees established for the proposed method.

Main Results:

• C-WOE significantly alleviates the adverse effects of ID samples in wild outliers.
• Demonstrated superior performance compared to state-of-the-art methods on various benchmarks.
• Validated the reliability of C-WOE in image processing applications.

Conclusions:

• C-WOE offers a simple yet effective approach to enhance OOD detection using readily available wild outliers.
• The reweighting strategy successfully suppresses negative supervision signals from ID samples.
• The method shows strong potential for real-world computer vision anomaly detection.