PWLR: Pairwise Witness Local Rejection for Boundary-Aware Out-of-Distribution Detection

2026-08-16

AI Summary

This study addresses the challenge of detecting near-distribution out-of-distribution (OOD) samples located proximate to class boundaries. We propose PWLR, a novel method that pioneers the use of language descriptions as explicit boundary evidence. By leveraging multimodal large models to generate and filter local visual cues, PWLR constructs reliable pairwise class verifiers and integrates global score calibration to achieve boundary-aware OOD detection. Extensive evaluations on benchmarks such as ImageNet-100 demonstrate that PWLR consistently outperforms existing strong vision-language baselines. The proposed approach effectively enhances detection robustness in boundary regions, establishing a new paradigm for semantics-guided OOD detection.

Abstract

Out-of-distribution (OOD) detection remains challenging for image classifiers, especially when near-OOD samples lie close to in-distribution (ID) class boundaries. Recent vision-language detectors improve OOD detection through class semantics, local prompting, or LLM-generated outlier concepts, but seldom use language as explicit boundary evidence between confusing ID classes. We propose Pairwise Witness Local Rejection (PWLR), which uses an MLLM offline to describe visible local cues that favor one ID class over a specific rival class. These cue phrases are then screened with ID-only data under a frozen vision-language backbone, so that only reliable local verifiers are kept. At inference, PWLR first retains a small set of globally plausible classes, then checks whether any of them is locally supported against its most relevant rivals, and finally combines this pairwise local evidence with the global class score through calibration. Experiments on ImageNet-100 far-OOD, cleaner/challenging OOD and near-OOD benchmarks show that PWLR consistently improves strong vision-language baselines across multiple backbones. Source code will be released.

Problem

• Out-of-distribution detection
• Near-OOD
• Boundary-aware
• Vision-language models

Innovation

• Pairwise Witness Local Rejection
• Boundary-Aware OOD Detection
• Multimodal LLM
• Local Visual Cues
• Vision-Language Models