Adapting Gaze Models to Individual Ocular Geometry

This direction is split into other directions: optical-visual axis offset. Its last version stays readable and citable here.

Version 1 of 153 papers70 conceptsfrozen 2026-10-04

Most modern gaze-tracking systems share a common frustration: they work remarkably well for the developers who built them, but often fail when deployed to a diverse general population. This failure isn't usually due to lighting or camera quality, but rather a fundamental biological reality. Every human eye is anatomically unique, and these subtle physical differences create a "personalization gap" that universal models cannot bridge.

The Visual Axis vs. The Optical Axis

To understand why gaze estimation is difficult, we must distinguish between where the eye is pointing and where the person is looking.

The optical axis is the line of symmetry of the eyeball itself. The visual axis, however, is the line connecting the fovea (the part of the retina responsible for sharp vision) to the object of interest. In humans, these two axes do not perfectly align. They are separated by a subject-specific angular offset known as the kappa angle.

As noted by Yiwei Bao and Feng Lu in their work on GazeGene, estimating the visual axis is significantly more challenging than the optical axis because the kappa angle is an internal physiological trait that cannot be directly seen in an image. Because this offset varies between individuals, a model trained on a general population will consistently produce systematic errors when applied to a new user.

The Anatomical Diversity of the Eye

Beyond the kappa angle, other structural variations introduce noise. Tsuyoshi Moriyama and Takeo Kanade demonstrated that human eye anatomy exhibits significant diversity in eyelid folds, iris dimensions, and lower-eyelid morphology Automated Individualization of Deformable Eye Region Model and Its Application to Eye Motion Analysis.

Even basic features like interpupillary distance (the distance between the eyes) vary by age and ethnicity, meaning a model calibrated for adults may fail when used by children Basic study on function of viewing distance measurement when using tablet computer. These differences mean that identical eye appearances in two different people can correspond to entirely different gaze directions.

Moving Beyond Tedious Calibration

The traditional solution to this problem is personal calibration: asking a user to stare at a grid of dots on a screen to "map" their unique eye geometry. However, this is tedious and often impractical for real-world use. Researchers are now pursuing three main strategies to eliminate or simplify this process:

1. Implicit and Self-Calibration

Instead of asking users to look at dots, systems can "steal" calibration data from natural behavior. Yoichi Sato and Yusuke Sugano proposed using mouse clicks as implicit gaze labels or leveraging visual saliency maps—probabilistic predictions of what a person is likely to look at in a video—to incrementally learn a user's eye parameters Gaze Sensing;towards calibration-free estimation. Similarly, Jixu Chen and Qiang Ji showed that combining 3D physical models with these saliency maps allows a system to reach high accuracy through natural viewing alone Probabilistic gaze estimation without active personal calibration.

2. Few-Shot Adaptation and Meta-Learning

If we cannot eliminate calibration, we can at least minimize it. The FAZE framework, developed by Seonwook Park and colleagues, uses meta-learning to adapt a model to a new user with as few as three images Few-Shot Adaptive Gaze Estimation. Another approach, SPAZE, treats personal anatomical differences as a tiny, 3-dimensional latent space, allowing the system to optimize just a few parameters per person to achieve significant accuracy gains Learning to Personalize in Appearance-Based Gaze Tracking.

3. Differential Estimation

Rather than predicting absolute coordinates, some models predict the difference in gaze between two images of the same person. Gang Liu and Yu Yu demonstrated that this differential approach effectively cancels out person-specific anatomical biases and environmental noise A Differential Approach for Gaze Estimation with Calibration.

Decoupling Appearance from Geometry

The most recent trend in the field is the move toward decoupled representation learning. In De^2Gaze, Yunfeng Xiao and Xiaowei Bai argue that models should explicitly separate permanent eyeball traits (like radius and fovea offset) from fleeting gaze movements. By isolating these identity-specific factors, models can become more robust to the physical variations of different users, eventually leading to gaze trackers that work "out of the box" for everyone.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adapting_gaze_models_to_individual_ocula_cfb4c5e2,
  title = {Adapting Gaze Models to Individual Ocular Geometry},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adapting-gaze-models-to-individual-ocular-geometry/rdr_cfb4c5e244ec465c84b282d755099994},
  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_adapting_gaze_models_to_individual_ocula_cfb4c5e2_v1,
  title = {Adapting Gaze Models to Individual Ocular Geometry},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adapting-gaze-models-to-individual-ocular-geometry/rdr_cfb4c5e244ec465c84b282d755099994/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}