Out-of-distribution Detection via Nearest Neighbors and Null Vectors on Principal Feature Space

Out-of-distribution (OOD) detection is crucial for ensuring the reliability of deep learning models in open-world environments, especially in safety-critical applications such as autonomous driving. Existing OOD detection methods typically focus on either classifier-based scores, which suffers from over-confidence, or distance-based approaches that lack fine-graininess due to their class-agnostic nature. Recently, some fused scoring functions have been proposed but be limited to single type of feature-based information. To address these challenges, we propose Nearest Neighbors with Null Space Analysis (k-NNuSA), a novel distance-focused approach that combines both distance to nearest neighbors within the in-distribution (ID) data and the residual against the ID principal subspace. Our method scales the generalized distance using classification confidence, enhancing fine-grained detection capability. We extensively evaluate <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k$</tex>-NNuSA on different CNN-based and transformer-based architectures, using ImageNet-1K as the ID reference set and a variety of OOD datasets, demonstrating its state-of-the-art performance of 96.23% AUROC and 19.34% FPR95. Our proposed scoring function consists of post-hoc operations which is instantly applicable without re-training.