Modeling the Biological Offset Between Optical and Visual Axes

36 concepts28 papers1 proposal

Imagine looking directly at a camera. A computer vision model analyzing your face can easily calculate the physical geometry of your eye: the boundaries of your iris, the curvature of your cornea, and the exact center of your pupil. This line of physical symmetry is known as the optical axis.

However, humans do not perceive the world along their optical axis. We see through the visual axis—the line connecting the object of interest to the fovea, a tiny, highly sensitive pit at the back of the retina. Crucially, the optical axis and the visual axis do not align. The angular discrepancy between them is a biological offset known as the kappa angle.

Because the fovea sits entirely inside the eyeball, its exact location is invisible from the outside. Worse, this offset varies significantly from person to person. This creates a fundamental bottleneck for universal gaze tracking systems: a model trained on a general population will exhibit a persistent, systematic "person-specific bias" when applied to a new user.

  Visual Axis (Line of Sight to Fovea)
         \
          \   Kappa Angle (θ)
           \=======> Optical Axis (Physical Centerline)
           |       |
         Fovea   Pupil

Why Universal Models Fail

When a generic deep learning model is trained to predict gaze coordinates from eye images, it attempts to learn a direct mapping from surface appearance to screen coordinates. However, individual structural diversity—including variations in eyelid folds, iris dimensions, and facial bone structure—means that the same physical gaze point results in wildly different visual appearances across individuals. As shown by Tsuyoshi Moriyama et al. in Automated Individualization of Deformable Eye Region Model and Its Application to Eye Motion Analysis and Yunyang Xiong et al. in Mixed Effects Neural Networks (MeNets) With Applications to Gaze Estimation, forcing a single global model to map these diverse inputs to a universal output space introduces a severe performance gap.

This issue is compounded by the unobservable nature of the kappa angle. As Seonwook Park et al. demonstrate in Towards End-to-End Video-Based Eye-Tracking, two users can exhibit identical physical eye orientations yet be looking at entirely different points on a screen due to their unique internal offsets. Mathematically, this introduces a contradiction into the training data: the network is presented with identical inputs mapped to different target outputs. Guanzhong Zeng et al. highlight in Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation that this physiological noise acts as a hidden domain shift, causing systematic angular errors across different datasets and users.

To resolve this, modern gaze estimation research focuses on isolating and modeling the person-specific bias. As Yiwei Bao et al. explain in GazeGene: Large-scale Synthetic Gaze Dataset with 3D Eyeball Annotations, estimating the optical axis is a straightforward geometric problem, but estimating the visual axis requires explicit calibration of individual-specific offsets.

Strategies for Personalization

To bridge the gap between the optical and visual axes without requiring tedious, multi-point user calibration, researchers have developed several distinct algorithmic frameworks.

1. Differential Tracking

Instead of predicting absolute gaze coordinates, a model can be designed to predict the difference in gaze direction between two frames of the same user. Gang Liu et al. introduce this concept in A Differential Approach for Gaze Estimation with Calibration. Because the biological kappa angle and facial structure remain constant for a given individual across frames, computing a spatial difference mathematically cancels out the static, person-specific bias.

2. Few-Shot Adaptation and Synthesis

Another strategy involves using meta-learning to rapidly adapt a generic model to a new user's anatomy using a minimal number of calibration points. Seonwook Park et al. propose a framework in Few-Shot Adaptive Gaze Estimation that disentangles gaze direction from personal appearance factors, allowing the model to calculate a user's specific bias vector from three or fewer samples.

To make the most of limited calibration data, Gang Liu et al. demonstrate in Improving Few-Shot User-Specific Gaze Adaptation via Gaze Redirection Synthesis that a system can synthetically warp a handful of user eye images to generate a dense, personalized training set. This maps out the user's unique eye geometry across various virtual gaze angles, effectively training out the person-specific bias.

3. Label-Free and Saliency-Driven Calibration

To eliminate active calibration altogether, systems can infer the kappa angle implicitly. Jixu Chen et al. show in Probabilistic gaze estimation without active personal calibration that by combining a 3D physical eye model with visual saliency maps—probabilistic predictions of where a user is likely looking on a screen—a system can incrementally learn the user's unique eye parameters during natural viewing.

Similarly, the authors of ELF-UA: Efficient Label-Free User Adaptation in Gaze Estimation utilize self-supervised meta-learning tasks on unlabeled user images to align the model's latent feature space with the individual's specific eye shape and visual axis, bypassing the need for explicit calibration targets.

Broader Biological Context

The challenge of individual biological variability extends beyond desktop eye tracking into broader computer vision and medical domains:

Ultimately, whether tracking human gaze on a screen, predicting driver head poses (Mahdi Rezaei et al. in Look at the Driver, Look at the Road: No Distraction! No Accident!), or optimizing stereoscopic video comfort (Miao Wang et al. in Comfort-driven disparity adjustment for stereoscopic video), accounting for individual biological deviations from the statistical norm remains the defining factor in achieving high-precision tracking.

Key Foundational Work

A Self-Calibrating, Camera-Based Eye Tracker for the Recording of Rodent Eye Movements (2010) — contributed 2 concept elements to this direction.
Automated Individualization of Deformable Eye Region Model and Its Application to Eye Motion Analysis (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_optical_visual_axis_offset_a17a1f5e,
  title = {optical-visual axis offset},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optical-visual-axis-offset/rdr_a17a1f5e52c34c92880989224d54b830},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_optical_visual_axis_offset_a17a1f5e_v1,
  title = {optical-visual axis offset},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optical-visual-axis-offset/rdr_a17a1f5e52c34c92880989224d54b830/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake