Automated Essential Concept Discovery for Few-Shot Out-of-Distribution Detection

Guangyao Chen, Kai Horstmann, Zhilong Wang, Fengqi You; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 3973-3983

Vision-Language Models (VLMs), trained on extensive internet-scale datasets, excel at identifying diverse objects and entities. Building on this foundation, several prompt learning methods have demonstrated effectiveness in Few-shot Out-of-Distribution (OOD) detection, which identifies OOD images from unseen classes during training with minimal labeled in-distribution (ID) examples. Despite their efficiency, these methods often lack clear interpretability. In this paper, we introduce a new framework, Automated Essential Concept Discovery (AECD), which leverages visual concepts for interpretable OOD detection. AECD consists of three stages: essential concept extraction, concept documentation, and essential attribute filtering. Initially, AECD extracts critical attributes from target categories using multimodal large language models, focusing on visual sample prototypes rather than individual image features. This ensures that the concepts reflect broad category traits. Subsequent stages involve using powerful language models to document and organize these concepts into a structured schema that captures and preserves category-specific visual attributes. Finally, a concept adapter refines and filters these attributes to determine the OOD status of target categories effectively. Experiments on the large-scale ImageNet OOD detection benchmarks demonstrate that our AECD method effectively enhances performance across both conventional and hard OOD detection settings.