Key Feature Replacement of In-Distribution Samples for Out-of-Distribution Detection

Out-of-distribution (OOD) detection can be used in deep learning-based applications to reject outlier samples from being unreliably classified by neural networks. Learning classify between OOD and in-distribution is difficult because data comprising the former extremely diverse. It has been observed that an auxiliary dataset most effective training a ``rejection'' network when its are semantically similar images. We first deduce images perceived they share common background, as networks incorrectly such with high confidence. then propose simple yet Key In-distribution feature Replacement BY inpainting (KIRBY) procedure constructs surrogate replacing class-discriminative features of marginal background features. The implemented using off-the-shelf vision algorithms, where each step within algorithm shown make increasingly data. Design choices studied extensively, exhaustive comparison state-of-the-art algorithms demonstrates KIRBY's competitiveness on various benchmarks.