Generalization-Enhanced Out-of-Distribution Detection via Explicit Spatial Alignment

Out-of-distribution (OOD) detection is essential for ensuring the reliability and safety of machine learning systems in real-world applications. A primary challenge in this task is distinguishing between in-distribution (ID) and OOD data while also generalizing to data that experience covariate shifts. Existing OOD detection methods based on implicit modeling have yet to effectively provide the precise feature distributions necessary for distinguishing between ID and OOD data and exhibit limited generalization capabilities when dealing with covariate-shifted in-distribution (CSID) data. In this study, we propose a simple and effective method named Explicit Spatial Alignment (ESA), which addresses this challenge through a novel explicit alignment strategy and a novel Weighted Feature Similarity Semantic Alignment (WFSSA) module. The explicit alignment strategy enhances the accurate modeling of ID training feature distributions by extracting inter-class discriminative features and clustering OOD features into specific regions using synthesized pseudo-OOD samples. The WFSSA module leverages ID data as a reference memory to emphasize consistencies and discrepancies between sample pairs, further improving generalization to CSID data. Extensive experiments across various benchmarks demonstrate that the proposed method significantly outperforms existing methods in discriminating between ID and OOD data while maintaining strong generalization across CSID data.