Adaptive Multi-Uncertainty Fusion for Robust Selective Classification under Heterogeneous Distribution Shifts

Out-of-distribution (OOD) detection and selective classification are critical for safety-critical computer vision systems, where models must reject both misclassified in-distribution samples (ID✗) and all out-of-distribution inputs simultaneously—a task formalized as Selective Classification with OOD Data (SCOD). Existing OOD detection approaches commonly improve ID/OOD separation at the expense of collapsing discriminability between correctly (ID✓) and incorrectly predicted ID data. The original Softmax Information Retaining Combination (SIRC) framework addressed this trade-off but suffers from manually fixed secondary confidence scores, static empirical hyperparameters, undifferentiated dual uncertainty, and poor generalization on modern vision architectures. This paper proposes SIRC-V2, an adaptive multi-uncertainty fusion framework with four core innovations: (1) A learnable Feature Gating Selector (FGS) that automatically mines multi-layer complementary confidence scores; (2) Bayesian Adaptive Parameter (BAP) optimization for dynamic hyperparameter tuning under varying ID/OOD mixing ratios; (3) Sparse Attention-weighted fusion (SA-SIRC++) that suppresses redundant score impact on ID✓/ID✗ separation; (4) Dual Uncertainty Decoupling (DUD) via Monte-Carlo feature perturbation separating aleatoric and epistemic uncertainty. Comprehensive experiments on classic CNNs, modern ConvNeXt/ViT architectures, 12 OOD datasets, plus real-world corrupted and adversarial benchmarks demonstrate SIRC-V2 consistently outperforms original SIRC and state-of-the-art baselines. Code will be released upon acceptance.