Enriching Category Representations with LLMs Towards Robust Zero-Shot OOD Detection

Abstract

Recent advancements in foundation models, particularly Visual-Language Models (VLMs), have enabled effective zero-shot Out-of-distribution (OOD) detection. Existing methods attempt to generate the names of OOD classes similar to in-distribution (ID) classes to explore the textual space of VLMs. However, they fail to integrate relevant ID information to reveal specific OOD features, thus limiting the distinction between ID and OOD classes. To address this issue, we propose a simple yet effective zero-shot OOD detection approach incorporating a specific semantic text generation strategy and a new regionally enhanced semantic OOD scoring function. In detail, we employ meticulously designed prompts to generate challenging OOD label texts using Large Language Models (LLMs). Subsequently, the specific semantic text generation strategy leverages LLMs to capture fine-grained textual representations of both ID and OOD classes. Additionally, the regionally enhanced semantic OOD score is formulated by adjusting the confidence of ID classes to improve OOD detection. Experiments demonstrate that our method achieves state-of-the-art (SOTA) performance on multiple OOD detection benchmarks. The code is available at repository .