GAFD-CC: Global-Aware Feature Decoupling with Confidence Calibration for OOD Detection

TL;DR

This work tackles out-of-distribution detection by bridging feature and logit information through a global-aware decoupling of features with respect to the full classification weight matrix. It introduces GAFD-CC, which first extracts positively and negatively correlated feature components via Delta f(x) = f(x) − μ_c and w_i = ∑_j W_{j,i}, computing directional scores ξ^+(x) and ξ^-(x), then fuses these with multi-scale logit confidences using Score(x) = ξ^+(x)/(s_{sample}+s_{class}) + ξ^-(x)/(s_{global}+s_{class}). The approach demonstrates strong OOD detection performance and robustness across datasets and backbones, supported by ablations, hyperparameter analyses, and cross-model evaluations. Practically, GAFD-CC provides a flexible, post-hoc method that improves reliability in open-world settings without additional training, by effectively leveraging global correlations between features and the classifier’s weight vectors. The results indicate meaningful gains over state-of-the-art baselines and solid generalization to hard OOD and near-OOD scenarios.

Abstract

Out-of-distribution (OOD) detection is paramount to ensuring the reliability and robustness of learning models in real-world applications. Existing post-hoc OOD detection methods detect OOD samples by leveraging their features and logits information without retraining. However, they often overlook the inherent correlation between features and logits, which is crucial for effective OOD detection. To address this limitation, we propose Global-Aware Feature Decoupling with Confidence Calibration (GAFD-CC). GAFD-CC aims to refine decision boundaries and increase discriminative performance. Firstly, it performs global-aware feature decoupling guided by classification weights. This involves aligning features with the direction of global classification weights to decouple them. From this, GAFD-CC extracts two types of critical information: positively correlated features that promote in-distribution (ID)/OOD boundary refinement and negatively correlated features that suppress false positives and tighten these boundaries. Secondly, it adaptively fuses these decoupled features with multi-scale logit-based confidence for comprehensive and robust OOD detection. Extensive experiments on large-scale benchmarks demonstrate GAFD-CC's competitive performance and strong generalization ability compared to those of state-of-the-art methods.

Figures (4)

• Figure 1: GAFD-CC Pipeline. Our proposed GAFD-CC framework involves two main steps: Global-Aware Feature Decoupling, which decouples features into positively/negatively correlated components to compute feature error scores ($\xi ^{+}(x), \xi ^{-}(x)$); and Confidence Calibration, which adaptively fuses logit-based scores ($s_{sample}, s_{class}, s_{global}$) at different scales with feature error scores to yield the final OOD detection score ($Score(x)$).
• Figure 2: Comparison of score distributions for ID (ImageNet-1K, blue) and OOD (SUN, purple) samples. (a, b) show the positive and negative feature scores from CADRef. (c, d) show the scores from our method. Our method demonstrates better separation (less overlap) between ID and OOD distributions.
• Figure 3: The impact of the hyperparameter $\lambda$ on OOD detection performance. This figure investigates how varying the hyperparameter $\lambda$, which governs the relative contribution of the two main components of the GAFD-CC score, affects the AUROC and FPR95 on test results.
• Figure 4: The impact of the hyperparameter $b$ on OOD detection performance. This figure investigates how varying the hyperparameter $b$, which governs the influence of class-related confidence on detection performance.