Answers are grounded in this paper and cite their sources.
Deep neural networks (DNNs) have become a fundamental component of modern visual computing systems, yet their reliability often degrades on out‐of‐distribution (OOD) inputs, i.e., test images that deviate from the training distribution. Such deviations may arise from high‐level semantic shifts, where images belong to unseen categories, or from image‐level covariate shifts, such as noise, blur, illumination changes, and compression artifacts. Although covariate shifts often preserve the original class identity, they can substantially distort low‐level visual statistics and alter model responses, making them difficult to detect with methods designed primarily for semantic novelty. This challenge becomes more severe when semantic and covariate shifts coexist, as they jointly affect different levels of visual abstraction. To address this issue, we propose AFDisc, a unified OOD detection framework that exploits the hierarchical representations of DNNs. AFDisc is motivated by the observation that shallow features are more sensitive to appearance variations, while deep features encode richer semantic information. Based on this complementarity, we introduce an attentive fusion module that adaptively integrates appearance‐sensitive cues from shallow layers and semantic cues from deep layers into a unified representation for detecting diverse distribution shifts. We further construct auxiliary supervision signals through controlled appearance perturbations, enabling the model to better capture the spectrum of real‐world visual degradations. Extensive experiments under semantic and covariate shift settings demonstrate AFDisc's effectiveness and its superior performance over prior state‐of‐the‐art methods.
No takes yet. Share an insight, caveat, or question.
Hou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: