Building webcam-based eye tracking without specialized hardware

Founder at @Decode by Entropik here. One of the things that surprises people most about MIRA is that we offer eye tracking: gaze, fixation, dwell time, heatmaps, attention of interest, using a standard consumer webcam. No Tobii. No IR sensors. No specialized hardware of any kind. 

The immediate question is always: how? 

The honest answer is that webcam-based eye tracking is a fundamentally different problem from hardware-based tracking. Hardware trackers use infrared illumination to precisely detect the pupil and corneal reflections. A webcam gives you RGB frames of a face, at whatever lighting conditions exist in the user's environment, at whatever angle they happen to be sitting at. 

The challenge isn't just detecting where the eyes are. It's detecting the gaze vector — the direction the person is actually looking, in a 2D image that varies with head position, lighting, distance from the screen, and the participant's facial geometry. Building a model that's robust to that variation took years of data collection and iteration. 

The number we report, up to 96% accuracy in controlled environments, is validated against data from physical eye trackers used as ground truth. In real-world conditions (variable lighting, participants not perfectly positioned), accuracy varies. We're transparent about that. But for the use cases that matter most in research, "did this participant look at this part of the screen?", "Where did their attention go during this product walkthrough?", Webcam eye tracking gives you a genuinely useful signal at a fraction of the cost and setup complexity. 

I'm curious what other builders have found when working at the intersection of computer vision and real-world environmental variability. What are the hardest constraints you've had to engineer around? 

And how do you communicate accuracy tradeoffs to customers without either overselling or underselling what the technology can do?