• Yitong Sun
• Zijian Zhou
• Cyriel Diels
• Ali Asadipour

Abstract

VR headsets often cause eye strain, yet eye-facing cameras only supply relative depth. We introduce a lightweight U-Net 3+ model that predicts metric 3-D periocular depth from a single monocular camera on any headset, enabling precise light-stimulus estimation. To sidestep scarce real data, we build a Dynamic Periocular Data Generation pipeline with UE MetaHuman, producing thousands of synthetic images from limited scans. Tested on 36 participants, the system delivers high global depth accuracy and trustworthy pupil-diameter readings, offering a practical basis for VR eye-health assessment.

https://doi.org/10.64560/32131242

Keywords