An Identity Level Protocol for Calibrated Open Set Face Recognition in Identity Scarce Scenarios

Open-set face recognition in unconstrained environments is hindered by identity scarcity, unreliable rejection of unseen individuals, and poorly calibrated decision confidence. This paper introduces an identity-level open-set evaluation protocol on the Labeled Faces in the Wild (LFW) dataset that explicitly models unknown identities during inference and enforces strict separation between training, gallery, and probe identities. The protocol enables joint assessment of open-set verification, identification accuracy, and decision calibration under both balanced and full-open operating conditions. Using this framework, a face-specific embedding model based on FaceNet with VGGFace2 pretraining is evaluated against strong generic and foundation-model baselines. Experimental results show that, using the proposed evaluation protocol, the best-performing face-specialized baseline (FaceNet trained on VGGFace2) achieves robust open-set discrimination and identification performance, with a ROC-AUC of 0.964, an equal error rate of 0.059, a Top-1 identification accuracy of 0.933, and an accepted-known identification accuracy of 0.982 under full-open evaluation. Furthermore, post-hoc calibration significantly improves confidence reliability without affecting discriminative performance. The proposed protocol and results demonstrate the importance of identity-aware evaluation and calibrated decision mechanisms for practical open-set face recognition.