RealEye Webcam v3 vs RedForestAI WebEyeTrack
Both RealEye Webcam v3 and RedForestAI WebEyeTrack run in the browser; here is how they compare, and how to read their accuracy numbers.
WebEyeTrack is an open-source research project from RedForestAI. RealEye Webcam v3 is the default eye-tracker for new RealEye studies. This article describes what each one is, where they are similar, where they differ, and why their published accuracy numbers cannot be compared directly. It does not declare a winner, because the two projects have never been tested against each other in a shared benchmark.
#What each eye-tracker is
#RealEye Webcam v3
RealEye Webcam v3 is the default webcam eye-tracker for new RealEye studies. RealEye Webcam v3 itself is built into the RealEye platform and is not published as a standalone open-source library, so you use it inside your RealEye studies rather than downloading and installing it separately.
- Technology: runs a machine-learning (neural network) model in the participant's browser inside a web worker. The model automatically falls back between execution backends as needed (WebGPU, then WebGL, then WASM) and uses MediaPipe for face detection.
- Calibration: a quick 4-point calibration (the four corners of the screen).
- Fielding devices: optimised for desktop computers (PCs and laptops). Selecting a smartphone or tablet as a fielding device automatically switches the study to RealEye Webcam v1.
- Hardware: needs nothing beyond a standard webcam.
You can read more about how v3 relates to the older RealEye Webcam v1 in RealEye Webcam v1 vs v3.
#RedForestAI WebEyeTrack
WebEyeTrack is an open-source, browser-native eye-tracking project published by RedForestAI in 2025. It is written to run in the browser with TensorFlow.js, and it also has a Python version for training and research.
- Licence: MIT for the code. The published model was trained on research-restricted datasets (MPIIFaceGaze, GazeCapture, EyeDiap), so the terms for the released model weights need to be checked before commercial use.
- Model: a small convolutional neural network called BlazeGaze (about 670 KB) that predicts the gaze point (x, y) directly.
- Calibration: on-device few-shot personalisation. The authors report that it needs up to 9 calibration samples.
- Head pose: model-based metric head pose, using MediaPipe 3D face reconstruction with radial procrustes analysis.
- Input: camera video only, plus the few-shot calibration samples; blink suppression uses the eye aspect ratio.
- Privacy: processing runs on the device, with no cloud dependency.
- Maturity: a young project (npm version 0.0.x), built by an academic team, with a small community. Because it is still young, anyone adopting it should benchmark it on their own task and participants before relying on it in production.
#How they are similar
- Both run the gaze estimation in the participant's browser instead of on a server.
- Both use a standard camera as the only input device, with no infrared or other dedicated eye-tracking hardware.
- Both use face analysis as the front end: RealEye Webcam v3 uses MediaPipe for face detection, and WebEyeTrack uses MediaPipe 3D face reconstruction for head pose.
- Both need a calibration step before gaze can be mapped to points on the screen.
- Both are built around ordinary webcam use rather than laboratory eye-tracking hardware.
#How they differ
| RealEye Webcam v3 | RedForestAI WebEyeTrack | |
|---|---|---|
| Maturity and support | The default eye-tracker for new RealEye studies, maintained as part of the commercial RealEye platform | A young open-source project (npm 0.0.x) from an academic team, with a small community |
| Licence and openness | Ships as part of the RealEye platform; this is the v3 engine, not a separate open-source library | MIT for the code; the released weights come from research-restricted datasets, so check their terms before commercial use |
| Calibration approach | Quick 4-point calibration (the four screen corners) | On-device few-shot calibration, reported by the authors as up to 9 samples |
| Head-pose handling | MediaPipe face detection, with the gaze model running in a browser web worker | Metric head pose from MediaPipe 3D face reconstruction with radial procrustes analysis |
| Reported accuracy | RealEye Webcam v3: median error 6.7% of the screen diagonal (v3 skin-tone validation). RealEye's separate all-tracker computer validation (not v3-specific) reports about 106 pixels on average | 2.32 cm point-of-gaze error on the GazeCapture benchmark, as reported by the authors |
#Reading the accuracy numbers honestly
The two published numbers are not directly comparable, for three reasons.
- They come from different benchmarks, and both are self-reported. The 2.32 cm figure is WebEyeTrack's own result on the GazeCapture dataset, measured by its authors. The 106-pixel figure is RealEye's own result from its validation study of real remote sessions. Neither number was produced by an independent comparison of the two systems.
- They are measured in different units. WebEyeTrack reports centimetres on a phone-sized screen; RealEye reports pixels. Accuracy figures are as reported by the authors on their own benchmarks and are not directly comparable across datasets: centimetres on mobile phones, degrees on desktop, pixels on a screen.
- They describe different setups. RealEye's computer figure is collected in uncontrolled, real-world remote conditions, where the participant's distance from the screen is not controlled and the screen size is self-reported. That is why RealEye reports accuracy in pixels: distance and screen size are needed to convert pixels into a visual angle or a physical measurement.
RealEye's own description of the 106-pixel figure is that the prediction is, on average, within about 1/20 of the screen's width and about 1/10 of the screen's height, with 72% of fixations landing within 150 pixels of the target centre, 80% within 200 pixels, and 91% within 250 pixels. RealEye also advises that Areas of Interest (AOIs) should cover approximately 5% of the width and 10% of the height of the displayed content so that the system's accuracy does not cause misclassification. See RealEye Accuracy on Computers for the full method and RealEye Webcam v3 Accuracy Across Skin Tones for the v3 validation data.
No head-to-head study of RealEye Webcam v3 and WebEyeTrack has been published, so this article does not say which one is more accurate.
#What this means for your study
- If you are running a study on RealEye, use the platform's own tracker. RealEye Webcam v3 is the default, it runs in the browser with a quick 4-point calibration, and RealEye publishes the measured accuracy and the matching AOI guidance.
- If you are building your own browser-based eye-tracker, WebEyeTrack is an open-source option worth evaluating. It is young, so benchmark it on your own task and participants before you rely on it.
- Do not put the two published accuracy numbers side by side as if they were measured the same way. They come from different devices, different benchmarks, and different units.
- For accuracy-sensitive work, follow the general guidance in How to get accurate results.
#Primary sources
- RedForestAI WebEyeTrack repository: https://github.com/RedForestAI/WebEyeTrack
- WebEyeTrack paper, WEBEYETRACK: Scalable Eye-Tracking for the Browser via On-Device Few-Shot Personalization (2025): https://arxiv.org/abs/2508.19544
- RealEye Technology Whitepaper