Slippage-Robust Gaze Tracking and Continuous Calibration

This direction is split into other directions: calibration drift. Its last version stays readable and citable here.

Version 1 of 134 papers41 conceptsfrozen 2026-10-04

Eye-tracking technology is fundamentally a mapping problem: how do we translate the pixels of an eye image into a coordinate on a screen or a point in 3D space? While modern systems can achieve remarkable precision, they are notoriously fragile. The moment a head-mounted display (HMD) slips a millimeter down a user’s nose, or a desktop user shifts their posture, the mathematical mapping breaks.

This phenomenon, known as slippage or calibration drift, is the primary barrier to seamless gaze interaction. Recent research has moved away from static, one-time calibration toward "set-it-and-forget-it" systems that use environmental context and behavioral patterns to self-correct in real-time.

The Fragility of Static Calibration

Most gaze trackers rely on an initial calibration phase where the user looks at specific targets. This establishes a fixed spatial relationship between the eye and the sensor. However, as Diederick C. Niehorster and colleagues demonstrate in The impact of slippage on the data quality of head-worn eye trackers, even minor actions like speaking or raising eyebrows can displace a headset enough to invalidate this mapping.

In mobile or high-intensity environments, such as sports, this problem is magnified. Bernardo R. Pires and Takeo Kanade found that wearable trackers in athletic settings frequently fail because they assume a constant eye-to-camera relationship. When the hardware shifts, the system "sees" the eye in a new position and incorrectly interprets this as a change in gaze direction.

Saliency as a Ground Truth Signal

One intuitive solution is to leverage the fact that humans do not look at the world randomly; we look at "interesting" things. By generating a saliency map of the user's environment, researchers can treat high-saliency regions as likely gaze targets.

Nicu Sebe and Roberto Valenti pioneered this "content-aware" approach in [What Are You Looking at?

  • Improving Visual Gaze Estimation by Saliency](/work/what-are-you-looking-at-improving-visual-gaze). If the eye tracker reports a gaze point in a blank area while a highly salient object is nearby, the system can infer a calibration offset and adjust itself. David Perra and Jan-Michael Frahm extended this to head-worn devices, using a local window of recent frames to continuously re-calibrate the system without any active user effort.

Reading Patterns for Vertical Drift Correction

Text provides an even more structured signal for auto-calibration. Reading follows a predictable "staircase" pattern: linear horizontal movement followed by a sharp "return sweep" to the start of the next line.

Systems like EyeO, developed by Akanksha Saran and John Langford, use "reading time" as a ground-truth signal. Because users look directly at text while reading but may look slightly off-target while typing, the system uses the reading phases to silently update the calibration. Similarly, the $RT^2H$ system by Sikai Yang and Wan Du uses these line-switching movements to specifically mitigate vertical drift, ensuring that gaze fixations stay aligned with the correct line of text even as the user's posture changes.

Geometric and Probabilistic Robustness

Beyond behavioral cues, researchers are improving the underlying geometric models of the eye to be inherently more "slippage-aware." Traditional 2D regression models are highly sensitive to sensor displacement. In contrast, 3D eyeball models—such as those proposed by Jiaxi Cao and Bin Li—attempt to estimate the eye's rotation center. By decoupling the gaze estimate from a fixed sensor frame, these systems can maintain accuracy even when the headset shifts.

Miika Toivanen and Kai Puolamäki further refined this by using Bayesian computer vision and adaptive filters. Instead of a single "best guess" for gaze, their system maintains a probability distribution that accounts for uncertainty caused by mechanical vibrations or movement. This allows the tracker to "wait" for high-confidence data before updating the calibration, preventing erratic jumps in the gaze point.

Why This Works

The shift from static to continuous calibration works because it transforms a hardware limitation into a software optimization problem. By treating the user's environment (saliency) and behavior (reading patterns) as a constant stream of "free" calibration points, these systems can:

  1. Eliminate Interruptions: Users no longer need to stop their task to look at calibration dots.
  2. Compensate for Physicality: They acknowledge that headsets will slip and users will move, building resilience into the algorithm rather than demanding perfect hardware stability.
  3. Improve Long-term Utility: By constantly correcting for drift, they enable long-term applications like gaze-controlled wheelchairs (Linus Lavi Raymond) or surgical navigation (Hong Chen) where manual recalibration is impossible.

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_slippage_robust_gaze_tracking_and_contin_09e67740,
  title = {Slippage-Robust Gaze Tracking and Continuous Calibration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/slippage-robust-gaze-tracking-and-continuous-calibration/rdr_09e67740b4554e3caa30f52e4ff2e963},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_slippage_robust_gaze_tracking_and_contin_09e67740_v1,
  title = {Slippage-Robust Gaze Tracking and Continuous Calibration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/slippage-robust-gaze-tracking-and-continuous-calibration/rdr_09e67740b4554e3caa30f52e4ff2e963/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}