Guides
What makes a good eye tracker? Part 2: Calibration-free operation
Neon measures where someone is looking, along with other information about their eyes, without requiring a calibration. This is made possible by NeonNet, the pipeline Neon uses to derive its measurements. Below, we explain the principles behind NeonNet, how it differs from conventional eye tracking, what optional personalisation is available, and what calibration-free eye tracking changes in practice.
Neon and NeonNet
For more than a decade, Pupil Labs has developed eye tracking methods for use in controlled and real-world environments. This work has included glint-free infrared imaging, models of how the eye works, and machine learning.
Each approach has different strengths. Models of eye geometry, physiology, and optics provide a physical basis for interpreting what the eye cameras see. Machine-learning models can learn the relationship between eye images and gaze across many different people.
NeonNet brings these strengths together in a single pipeline. Its interpretation of each image draws both on patterns learned from data and on models of the eye. Together, these approaches allow NeonNet to derive physically meaningful measurements from eye images and help it maintain performance across diverse wearers and challenging recording conditions.¹,²
The measurements provided by NeonNet include gaze direction and the 3D state of each eye. Gaze direction is measured in relation to the forward-facing scene camera. The 3D eye state is calculated separately for each eye. It includes the position and orientation of the eyeball, together known as its 3D eye pose, and physical pupil diameter in millimetres.²
How NeonNet differs from conventional eye tracking
Conventional eye tracking systems look for discrete visual features such as the pupil and, in some cases, glints, which are reflections of an infrared light source on the cornea. They then require a calibration procedure, which converts these tracked features into a gaze direction.
To calibrate, the wearer is typically asked to look at one or a series of targets at known positions. The system then fits a mapping between the tracked eye features and the point being viewed to yield a gaze measurement.
There are different ways to create this mapping. Some methods fit a mathematical function directly, based on the position or shape of the pupil in the eye image. Others build a model of the eye's geometry and optics, then fit that model to the calibration data.
The key differences are what information the systems use and when the relationship is established. Conventional systems depend on explicitly tracked features and learn part of the relationship from each wearer during calibration. Their mapping can therefore be affected if a required feature is lost or the eye tracker shifts relative to the wearer.
NeonNet instead works from the overall eye image and applies a relationship learned across many wearers. It does not depend on any single feature, such as a corneal glint, or need to fit a new gaze mapping each time someone puts on the eye tracker.
Optional personalisation
Although Neon does not require calibration, its gaze measurements can still be adjusted for a particular wearer. For applications that need a smaller systematic gaze error, Neon provides optional offset correction.
Offset correction shifts Neon’s gaze measurements by a consistent amount and direction. It can account for person-specific physiological differences, as well as other factors that cause a wearer’s gaze to differ consistently from the population average.¹
Offset correction is not the same as calibration. It does not create Neon’s gaze mapping or require a fixed routine at the start of the session. It can also be calculated retrospectively from a recording. The Accuracy article explains how NeonNet’s accuracy was evaluated and how offset correction affected the results.
What calibration-free eye tracking changes in practice
For research, calibration-free operation reduces setup between participants and sessions. It also removes the instructed fixation task normally required before recording. This can extend eye tracking to some participants who cannot reliably complete a conventional calibration, including infants and young children, people with low vision or nystagmus, and people who find the task difficult because of a cognitive or clinical condition.
For industry, calibration-free operation lets a worker put on Neon and begin their task without completing a setup procedure. If another worker takes over, the device can simply be passed on and used straight away. An eye-tracking specialist does not need to be present to run or repeat calibrations.
For integrators, calibration-free operation removes calibration logic from the surrounding application. The application does not have to display targets, assess the result, store a calibration profile, or decide when to repeat the process.
Together, these practical changes can enable research and applications that would otherwise be impractical or require substantial resources.
From calibration-free operation to robustness
Calibration-free operation is one part of what makes Neon robust. Because NeonNet does not depend on a wearer-specific gaze mapping, there is no calibration for a change in headset position to invalidate. Its use of the overall eye image, rather than any single feature such as a corneal glint, also helps it handle changes in what the cameras see. The next article examines these broader aspects of robustness.
Where to go next
This article is part of a series exploring what makes a good eye tracker. Related topics include:
¹ Baumann & Dierkes, Neon Accuracy Test Report, Pupil Labs, 2026. doi:10.5281/zenodo.18504792
² Pfeffer & Dierkes, Neon Pupillometry Test Report, Pupil Labs, 2024. doi:10.5281/zenodo.10057185
See how NeonNet measures gaze without calibration.
See what researchers have published using calibration-free eye tracking.