Enhancing out-of-distribution detection via local-image set.

Out-of-distribution (OOD) detection is essential for ensuring the reliability and safety of machine learning systems in open-world scenarios. Most existing methods rely on global representations of the entire image to determine whether a sample is OOD, assuming that the most informative features are uniformly distributed or sufficiently captured at a coarse spatial level. A common yet underexplored failure case arises when in-distribution (ID) images contain small foreground objects and large background regions, which often leads to high false positive rates (FPRs) across various visual recognition models. We attribute this issue to inherent spatial biases, scale sensitivity, and limitations in global feature extraction that limit clear foreground-background separation. To tackle this challenge, we propose a simple yet effective selection-then-calculation strategy, which generates multiple randomly local views of an input image and selects the one with the lowest predictive entropy for final OOD scoring. This process implicitly emphasizes salient foreground regions while suppressing irrelevant background noise, thereby improving the ability to distinguish between ID and OOD inputs. Our method is model-agnostic and can be seamlessly integrated into existing OOD detection pipelines as a plug-and-play component. Experiments demonstrate consistent improvements in OOD detection performance across various architectures.