13 March 2024 in European Journal of Wood and Wood Products
11 May 2026 in IEEE Access
Estimating ocular alignment in unconstrained settings faces challenges due to Noise-Induced Binocular Inconsistency (NIBI) which results from random events like blinking, head motion and tracking noise leading to temporary inter-ocular inconsistencies. Existing solutions typically follow either a framebased or a classification-based pipeline which often fail to yield reliable results especially when required to produce continuous measurements for clinical interpretation. In this work, we introduce a calibration-free deep multi-expert architecture that is capable of estimating the required prism diopters in a matter of seconds, by analyzing a short monocular video clip in real-time. Unlike prior methods, we tackle ocular alignment as a temporal problem, which results in improved robustness. Our approach consists of a Geometric Expert (Living Ruler) which outputs iris-normalized and scale invariant features. These are followed by a Trajectory Expert that encodes both eye movement temporal dynamics and deep image features, which are then passed through a Stability Supervisor. The Supervisor — a fully convolutional Transformer encoder — outputs frame-wise reliability scores which are subsequently used to suppress any noise-induced inconsistencies. Due to the lack of publicly available datasets with corresponding prism diopter annotations, we simulate geometry-based reference deviation values for eyes with real human eye geometry from the Columbia Gaze Dataset. Experimental results demonstrate that the proposed method outperforms other solutions in terms of both accuracy (MAE) of 1.3038Δ and R2 of 0.9963 for the range 0–50Δ and stability where a temporal fusion strategy leads to better results compared to frame-wise approaches especially when affected by transient noise. These findings indicate that continuous, calibration-free estimation of ocular deviation from standard video input is feasible, supporting potential applications in tele-screening and longitudinal monitoring.
7 September 2026 in SSRN Electronic Journal
Few-shot gaze personalization involves adapting a pretrained estimator to a new user using a limited number of calibration frames, with the assumption that each frame accurately captures the user looking at a known target. While image corruption that violates this assumption can be identified from the pixels, an attention lapse cannot, as it results in a clean image paired with an incorrect gaze label. Label noise has been extensively studied at the scale of training sets, where population statistics are available, and in few-shot classification, where prototypes exist; however, few-shot regression lacks both. We investigate the detectability of a corrupted calibration label from nine frames. By scoring each frame based on the personalized model's prediction residual, we locate a detectability floor: an ROC-AUC of 0.7 is achieved at a label displacement of δ = 4.2–4.8° against a personalized error of 3.57°, which remains stable across support-set contamination ranging from 10% to 40%. This floor is model-relative, with the crossing correlating with a subject's own error at r = +0.79, allowing the floor of a model with known accuracy to be anticipated rather than remeasured. Compared to five baselines, the separation is determined by mechanism rather than provenance: population-statistical scorers offer no advantage at nine frames, while per-frame residuals transfer intact, including the small-loss criterion that prior work explicitly excludes. We recommend using the one-adaptation resubstitution residual at a threshold of 3.42°, which recovers 1.3 corrupted frames per nine-frame session while incorrectly flagging 0.5 clean ones. Under a structured model of attention lapse, 8–10% of corruptions fall below the floor and are inherently undetectable, increasing to 16% for the most affected subject. Decomposing the damage a corrupted frame does, we find that label corruption exceeds frame loss only above the floor, while image corruption never does.
1 January 2026 in IEEE Access
1 January 2026 in IEEE Access