Eye tracking, without the eye tracker
Eye tracking has always meant buying a dedicated sensor. We are removing it. Our gaze engine turns the cameras already sitting in laptops, kiosks, and machines into an input you can build on.
The problem we work on
Detecting eyes is solved. Keeping gaze useful for hours is not.
Almost anything can track a gaze for ten seconds. Then the person shifts, the light moves, the laptop is a different one, and hour two looks nothing like the calibration. That gap, not the demo, is what keeps eye tracking out of shipping products. We close it.
1.20° of gaze error seated, once the engine has adapted to you. The best known webcam system measures 1.64°.
| Condition | llocus | Reference |
|---|---|---|
| Seated, adapted | 1.20° | 1.64° |
| Moving, adapted | 1.78° | — |
| Seated, cold | 1.65° | 1.64° |
| Precision | 0.72° | 0.85° |
Gaze error is the angle between where someone looked and where the engine said they did, so smaller is better. Precision is how tightly repeated samples of the same point cluster. Adapted means after 20 activations of ordinary use; cold means the shipping configuration before any adaptation. Reference is Tobii Nexus, working from the same class of camera. Measured across 15 hardware setups in uncontrolled rooms and changing lighting, not a lab bench. Dedicated eye tracking sensors are a separate category and can still be more accurate, but they need a sensor bolted to the product.
Camera frames go in. A gaze signal comes out. Calibration, per-user adaptation, and movement handling live inside it instead of on your roadmap.
Fitted to each user, still learning as it runs, on about one CPU core.
No two faces, cameras, or seating positions produce the same picture, so we do not ship one frozen model and hope. A short calibration fine-tunes the engine to the individual user, on the machine they are actually using.
Real use tells the engine when it was probably wrong. We feed that back in to keep improving during normal use, instead of sending people back to a calibration grid every time something drifts.
Capture, inference, and smoothing together. No GPU budget, no dedicated silicon, and enough headroom left over for the product the gaze signal is feeding.
Gaze is computed on the device holding the camera. No frames leave it, and no llocus server sits in the loop.
Gaze is computed on the machine holding the camera. Webcam video does not travel anywhere to be turned into a gaze signal, because the work happens where the camera already is.
The engine does not need us to be reachable to do its job. Nothing about a user, their face, or where they looked has to reach us for the product around it to work.
For an embedded or industrial deployment, camera data staying on the device is usually the difference between a short conversation with a security team and a long one.
Where this goes
Why we also ship an app
llocus desktop is a real product and the engine's test surface at the same time.
No lab produces a thousand different webcams, faces, rooms and postures. People do. llocus desktop lets anyone read and navigate with their eyes, and every hour of that use shows us where the engine is still wrong, before the same engine goes into somebody else's product.
See llocus desktopCommercial inquiries
Microphone, camera, keyboard, touch. Each one is just a signal software can ask for. Eye tracking is the last one still locked behind special hardware, and that is the whole reason we exist.
Tell us what you want to point a camera at: the device, the conditions it runs in, and what a gaze signal would let you do. You will get a straight answer about fit from the people building the engine.
inquire@llocus.cc