COMBOOD: A Semiparametric Approach for Detecting Out-of-distribution Data for Image Classification
This work addresses the challenge of effectively detecting near-distribution out-of-distribution (near-OOD) samples in image classification inference, a task where existing methods often fall short. To this end, the authors propose COMBOOD, an unsupervised semi-parametric framework that uniquely integrates non-parametric nearest-neighbor distances with parametric Mahalanobis distances in the feature embedding space to produce a unified confidence score. This fusion enables robust performance across both near-OOD and far-OOD scenarios. COMBOOD is compatible with diverse feature extractors and exhibits computational complexity that scales linearly with the embedding dimensionality. Extensive evaluations on OpenOOD v1/v1.5 benchmarks and document datasets demonstrate that COMBOOD consistently outperforms current state-of-the-art methods, with most improvements achieving statistical significance.