Why Feature Magnitude Deceives OOD Detectors: An Angular Separation Perspective
Published in European Conference on Computer Vision (ECCV), 2026
This paper investigates why feature magnitude deceives out-of-distribution (OOD) detectors and proposes a principled angular separation perspective to address the limitations of magnitude-based OOD scoring. By decoupling feature magnitude and direction, the method achieves more robust OOD detection performance.
Recommended citation: Hanlin Li, Jing Ma, Zehang Wei, Jiamin Yan, Xiang Xiang. (2026). "Why Feature Magnitude Deceives OOD Detectors: An Angular Separation Perspective." ECCV 2026.
Download Paper