Abstract

Estimating ocular alignment in unconstrained settings remains challenging because Noise-Induced Binocular Inconsistency (NIBI), caused by blinking, head motion, and tracking noise, can produce temporary inter-ocular inconsistencies. Existing frame-based or classification-based methods often cannot provide reliable continuous measurements for clinical use. In this study, we introduce a calibration-free deep multi-expert architecture for real-time prism diopter estimation from short monocular video sequences. Unlike approaches that treat ocular alignment as a static image-, feature-, or frame-wise task, the proposed framework formulates it as a temporal regression problem to better handle dynamic image variations. The framework integrates three complementary components: a Geometric Expert, termed the Living Ruler, which extracts iris-normalized and scale-invariant features; a Trajectory Expert, which encodes temporal eye movement dynamics together with deep image features; and a Transformer-based Stability Supervisor, which estimates frame-wise reliability scores for reliability-aware temporal fusion by emphasizing stable observations and suppressing unreliable ones. Because publicly available datasets with prism diopter annotations are limited, geometry-based reference deviations were simulated using the real human eye geometry derived from the Columbia Gaze Dataset. Experimental results show that the proposed method achieves high accuracy, with a mean absolute error of 1.3038\Delta $ and an R2 of 0.9963 over the range of 0-50Δ , while maintaining strong temporal stability. Evaluation on real-world data without fine-tuning further demonstrated generalization across appearance, head pose, and imaging variations. These findings suggest that continuous calibration-free ocular deviation estimation from standard video input is feasible for tele-screening and longitudinal monitoring.