Protodetect: Prototype Compactness and Inference Calibration for Few-Shot Out-of-Distribution Detection

Abstract

Detecting out-of-distribution (OOD) samples from limited labeled data is important for reliable recognition under semantic novelty. Existing few-shot approaches use synthetic or auxiliary unknowns, model only in-distribution (ID) data, or rely on large pretrained vision–language models. This paper presents Protodetect, a metric-learning framework for the conventional episodic, prototype-based setting. Model weights are optimized using ID episodes only, while held-out OOD samples during meta-validation are used to select inference hyperparameters. Protodetect combines the prototypical classification objective with a prototype-anchored triplet loss to encourage compact within-class representations and separation between known classes. At inference, temperature scaling and gradient-based input preprocessing are applied to the prototype-distance scores. Experiments on miniImageNet and tieredImageNet evaluate both closed-set accuracy and OOD AUROC. Across three independent training seeds, Protodetect obtains a mean AUROC of 81.13% under the reported 5-way 5-shot miniImageNet protocol, 1.28 percentage points above the strongest evaluated baseline. On tieredImageNet, the corresponding mean AUROCs are 75.74% and 83.65% in the 1-shot and 5-shot settings, respectively. These results establish Protodetect as an effective incremental approach under the evaluated episodic protocols.