Knowledge Base

Tracking Technology

You can use EyeReader in a flexible and mobile lab environment or within the FaceReader Online platform for online testing.

How It Works

The EyeReader algorithm estimates the gaze direction and relates that to an image on the screen. The first step is face analysis to locate key-points in the face (such as eyes, nose, etc.) and estimate the head pose.

The eye image and head pose information are then used by a pre-trained neural network (a convolutional neural network or CNN) to produce a 3D gaze vector (a line segment in three directions, x,y,z) representing the direction of the person's eye gaze. The neural network is trained with many labelled datasets of screen locations and video recordings to learn the relation between the image of the eyes and the gaze vector.

To go from a 3D model of the gaze to a location on a screen one needs to know exactly where the screen is. With the help of a calibration task in which the participant follows dots on a screen, the 3D vector can be mathematically decoded to 2D gaze x,y points on the screen.

EyeReader

We have tested on our own validation dataset, which includes data from 17 participants (12 males/5 females, 9 aged 18-29/8 aged 30-59) with EyeReader and a Tobii Nano eye tracker with natural lighting conditions. Participants performed a calibration task and a few psychological tasks. They were situated with a distance of approximately 75 cm in front of a 1920x1080 23-inch screen and also performed a calibration task on a 15-inch laptop (without eye tracker).

The results on the 32-dot calibration task (different dots used for calibration and testing) on a desktop computer show that the system predicts gaze points on screen with an average deviation of 2.4 ±0.6 cm (or an average of 113.6 pixels, and average of 5.2% deviation of the screen).

Also, the results demonstrate that a large amount of sections of the screen can be discerned with high accuracy (it can estimate a gaze location in 1/16th of the screen with 96% accuracy), albeit not in the range of hardware eye tracking systems (having over 99% accuracy for 1/16th of the screen). If one wants to study a difference in fixation between two points on the screen, it is ideal if they are in different 1/16th sections of the screen.

The results are comparable with competitor systems on the market, although it should be noted that direct comparison is not possible due to different screen sizes (making noting errors in pixels problematic) and different testing settings.

Laptop Versus Desktop

Another outcome of this test was that the system is more accurate on the laptop (Mean Absolute Error: 2.3 ±0.5 cm) in comparison with the desktop (MAE: 2.5 ±0.7 cm). This is likely due to the smaller screen and the fixed ideal alignment of the in-built webcam of the laptop.

The results are similar for brown (MAE: 2.5 cm) in comparison to blue eyes (MAE: 2.4 cm), and for those not wearing glasses (MAE: 2.4 cm) in comparison to those that are wearing glasses (MAE: 2.7 cm). However, this is based on only 3 people who are wearing glasses (with thin lenses); when there is reflection of the surrounding light this will likely cause issues.

Validation of This Application

An important but simple finding within eye tracking research is that people look longer at what they find interesting. Participants performed two simple choice tasks, where they viewed a few images of items (e.g. cookies, chocolate, etc.) and were allowed to choose one afterwards.

A comparison of the EyeReader gaze estimations with the eye tracker estimations shows a high correlation between the X and Y coordinate estimation (X: Mean r = 0.87, Y: Mean r = 0.77).

Gaze estimations can be classified into being a saccade (fast movements with less visual information processing) or fixation (slower eye focus relevant for information processing) based on the speed of the gaze movements. For each category that was presented, total duration of fixation time was calculated.

Both systems resulted in the same interpretation of the results: 79% of participants looked longest at the category that they preferred (choice as a reward for participation), with EyeReader giving an average of 9.3 seconds on the preferred category.

Source: EthoVision XT 18 - THC - Trial and Hardware Control, Noldus Information Technology