Iris recognition on head-mounted displays (HMDs) is gaining attention for securing Extended Reality (XR) systems, but practical HMD imagery is low-resolution, off-axis, and prone to specular highlights. We present a 1:1 verification pipeline for periocular/iris images acquired by an AR/HMD device that combines deterministic segmentation with deep metric learning. Classical computer vision is used to preprocess the eye region, detect pupil and iris ellipses, and unwrap the iris into a normalized strip, which is then fed to a ResNet-50 backbone with a 128-dimensional embedding head trained using triplet loss. Verification is performed by thresholding the Euclidean distance between L2-normalized embeddings. On a subset of the public PolyU AR/VR iris dataset comprising 92 subjects, we evaluate under a balanced one-session protocol with 2,000 genuine and 2,000 impostor pairs. The proposed verifier achieves ROC–AUC = 0.9803 and EER = 6.40%. At a Youden operating point, accuracy and balanced accuracy are 94.03% with FAR = 7.55% and FRR = 4.40%, while at a looser security budget of FAR = 10% the system attains TAR = 97.20%. These results show that a simple metric-learning encoder coupled with deterministic segmentation provides a strong, transparent baseline for iris verification in HMD-based XR authentication.
Awadallah et al. (Thu,) studied this question.