TINS: Test-time ID-prototype-separated Negative Semantics Learning for OOD Detection

2026-05-11

AI Summary

Existing out-of-distribution (OOD) detection methods relying on static negative labels struggle to capture the dynamic and diverse nature of OOD concepts and are prone to in-distribution (ID) sample contamination during test-time adaptation. This work proposes a novel paradigm that dynamically learns sample-specific negative semantics at test time by inverting image-to-text modalities to generate negative semantic representations. To ensure these negatives remain disentangled from ID semantics, the method incorporates an ID prototype separation regularizer. Furthermore, a grouped aggregation scoring mechanism combined with a dynamic buffer update strategy effectively mitigates ID contamination while enhancing OOD coverage. Evaluated on benchmarks including Four-OOD, the approach significantly outperforms strong baselines; notably, when using ImageNet-1K as the ID dataset, it reduces the false positive rate at 95% true positive rate (FPR95) from 14.04% to 6.72%.

Abstract

Vision-language models enable OOD detection by comparing image alignment with ID labels and negative semantics. Existing negative-label-based methods mainly rely on static negative labels constructed before inference, limiting their ability to cover diverse and evolving OOD concepts. Although test-time expansion provides a natural solution, naively learning negative semantics from potential OOD samples may introduce hard ID contamination. To address this issue, we propose a \textbf{T}est-time \textbf{I}D-prototype-separated \textbf{N}egative \textbf{S}emantics learning method, termed \textbf{TINS}. TINS learns sample-specific negative text embeddings via image-to-text modality inversion and introduces ID-prototype-separated regularization to keep them separated from ID semantics. To further stabilize negative semantics expansion, TINS employs group-wise aggregation scoring and a buffer update strategy. Extensive experiments across Four-OOD, OpenOOD, Temporal-shift, and Various ID settings show consistent improvements over strong baselines. Notably, on the Four-OOD benchmark with ImageNet-1K as ID, TINS reduces the average FPR95 from 14.04\% to 6.72\%. Our code is available at https://github.com/zxk1212/tins.

Problem

• OOD detection
• negative semantics
• test-time learning
• ID contamination
• vision-language models

Innovation

• test-time learning
• negative semantics
• ID-prototype separation
• OOD detection
• vision-language models