Antoine Boniface | AlphaLum: Can a glasses platform classify eye primitives such as saccades and smooth pursuits with high accuracy?
07:36 - 09:51
Other snippets from this talk
Summary of the clip:
Can a glasses platform classify eye primitives such as saccades and smooth pursuits with high accuracy?
The system was adapted to one specific application: classification of eye primitives. These are the simple building blocks the eyes can perform: saccades, smooth pursuits, fixation and blinks. Those four movements form the basic set; other eye movements are compounds of them. The demonstration uses the full platform to predict which primitive the user is performing, moving beyond raw sensing to labelled eye-movement classes.
The measured capabilities are explicit. Eye speed can be tracked up to 500 degrees per second. At 250 frames per second, resolution is around two degrees per second. The eye box is also huge, which is credited to the optical combiner that sits on the lens. These figures matter because they define the dynamic range and angular precision available for classifying fast saccades, slow pursuits and stable fixations.
In the demonstration, a user wears glasses while dots or colours on a screen emulate different movements: the middle dot is fixation, the yellow one smooth pursuit, and another saccades. The user follows one dot, and the full platform predicts the action. On the right, the background colour is ground truth, and the confidence curve matches it well. For a saccade, prediction accuracy was around 80 to 90 percent.
In this short video, you can learn:
* Eye primitives are saccades, smooth pursuits, fixation, and blinks; other movements are compounds.
* The platform tracks eye speed up to 500 degrees per second and resolves about two degrees per second at 250 frames per second.
* In the demo, saccade classification reached roughly 80-90% accuracy, with confidence curves matching ground-truth background colours.
📋 **Clip Abstract** The clip shows a glasses-based platform deriving a task kit for classifying eye primitives: saccades, smooth pursuits, fixation and blinks. A demonstration with dots or colours as targets reports up to 500 degrees per second speed, about two degrees per second resolution at 250 frames per second, a huge eye box from an optical combiner on the lens, and roughly 80-90% saccade prediction accuracy.
About the speaker:
* Speaker: Antoine Boniface
* Company: AlphaLum
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#EyePrimitiveClassification, #SaccadeDetection, #SmoothPursuitTracking, #OpticalCombiner, #EyeTracking, #WearableSensing
This is a highlight of the presentation:
Camera-Free, Low-Power Eye Sensing for AI Glasses Using Self-Mixing Interferometry
More Highlights from the same talk.
03:11 - 05:26
How can a single VCSEL and photodiode replace a camera for low-power eye tracking?
How can a single VCSEL and photodiode replace a camera for low-power eye tracking?
SMI stands for self-mixing interferometry, a deterministic physical phenomenon that occurs when light reflects back into a laser cavity. The laser used here is a VCSEL. VCSELs are not very good or stable lasers: their cavity is super small and they are super sensitive to external feedback. With a target in front, any light re-entering the cavity perturbs the system, and output power is modulated by the target's motion.
Measurement is easy: because VCSELs are two-side emitted, a photodiode placed in the stack at the back captures the signal. The SMI module measures in the time domain, not spatial resolution like a camera. It produces a time trace of fringes from interferometry, yielding velocity and distance. Precision is millimeter per second, can go up to kilohertz, but is limited to 1,000 hertz here.
Power goes down to five milliwatt, and depending on the application, below two milliwatts. The sensor is total ambient light immune because it is based on interferometry: it is sensitive to its own light, so ambient light does not matter. It operates over different surfaces and works at short or long range; for eye sensing, range is up to a few centimeters. The SMI module is a very important piece.
In this short video, you can learn:
* Self-mixing interferometry uses light reflected into a VCSEL cavity, where target motion modulates output power deterministically.
* A photodiode at the back of the two-side-emitting VCSEL captures a time-domain fringe trace, not camera-like spatial resolution.
* The SMI module measures velocity and distance at millimeter-per-second precision, runs down to five milliwatts or below two milliwatts, and is immune to ambient light.
📋 **Clip Abstract** Self-mixing interferometry uses a VCSEL whose output power is modulated when reflected light re-enters its cavity, allowing a back-side photodiode to capture target motion as time-domain fringes. For eye sensing, this SMI module provides millimeter-per-second velocity and distance over a few centimeters, with five-milliwatt or below-two-milliwatt power and immunity to ambient light.
About the speaker:
* Speaker: Antoine Boniface
* Company: AlphaLum
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#SelfMixingInterferometry, #VCSEL, #Photodiode, #TimeDomainFringes, #EyeTracking, #ARGlasses
16:30 - 17:49
Why does calibration-free eye sensing tolerate glasses slippage, with no coordinate system and only eye speed?
Why does calibration-free eye sensing tolerate glasses slippage, with no coordinate system and only eye speed?
The key distinction is eye sensing versus gaze tracking. Gaze needs a reference coordinate system to define an angle with respect to something, which is why calibration is normally required. Here the system measures only the speed of the eye, so it does not need to reference that measurement to a fixed coordinate system. That removes the calibration step when glasses move on the head.
The eye must sit inside the eye box. The optical system provides a very large eye box, so small shifts from glasses slippage are tolerated. Because the measurement is speed, not absolute position, there is no need to recalibrate after the glasses move. The approach works at 940 nanometers, which is invisible to the eye.
This design changes what the system must guarantee: not an absolute gaze coordinate, but that the eye stays in the eye box. The large eye box provides that tolerance. Measuring speed rather than position means movement of the glasses does not create a calibration error. The 940-nanometer wavelength keeps the sensing illumination invisible, supporting use in AI glasses.
In this short video, you can learn:
* Eye sensing measures only the speed of the eye, so it avoids the coordinate-system reference that makes gaze tracking require calibration.
* A very large eye box keeps the eye within the optical system even when the glasses slip on the head.
* Operation at 940 nanometers keeps the sensing illumination invisible to the eye.
📋 **Clip Abstract** Calibration-free operation comes from measuring eye speed rather than gaze angle, so no absolute reference coordinate system is required. A very large eye box tolerates glasses slippage, and 940-nanometer operation keeps the sensing illumination invisible to the eye.
About the speaker:
* Speaker: Antoine Boniface
* Company: AlphaLum
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#CalibrationFreeEyeSensing, #LargeEyeBoxOptics, #940nmEyeSensing, #EyeSpeedMeasurement, #AROptics, #SmartGlasses
10:56 - 12:09
What is the power breakdown of a camera-free eye-tracking system operating under two milliwatts?
What is the power breakdown of a camera-free eye-tracking system operating under two milliwatts?
The system can go below five milliwatts; in this specific test it goes below two milliwatts. Most power is used for driving the voxels and the readouts. The rest is quite small. Compute takes 0.4 milliwatt, covering prediction after raw-signal capture. The compute can be decreased by a few milliwatts by choosing the model. That is the breakdown.
You do not need a huge data center to train the model. It is very small, and the goal is to have this like a GI, basically directly on the device. You can do it because you do not need a huge architecture to run the platform and do the prediction. So the model and platform can run predictions locally.
The system can predict the primitives. That gives information. The speaker asks what you want to do with that, and whether you can do more than this. That question frames the next step: not only capturing raw signal and predicting primitives, but using that information for additional action. The clip ends on that open technical challenge, after establishing the low-power capture and prediction path.
In this short video, you can learn:
* In this specific test, the system operates below two milliwatts, though it can go below five milliwatts.
* Most power goes to driving voxels and readouts, while compute uses 0.4 milliwatt for prediction after raw-signal capture.
* The model is very small and does not need a huge data center or huge architecture, aiming to run directly on the device like a GI.
📋 **Clip Abstract** The clip details a power budget below two milliwatts in this specific test, with most power used for driving voxels and readouts and 0.4 milliwatt for compute. It stresses that the prediction model is small enough to avoid a huge data center and potentially run directly on the device, while asking what more can be done with the predicted primitives.
About the speaker:
* Speaker: Antoine Boniface
* Company: AlphaLum
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#VoxelReadoutPower, #SubTwoMilliwatt, #RawSignalCapture, #OnDeviceModel, #SmartGlasses, #EdgeAI