SaccadeJS: Webcam eye tracking for behavioral experiments in a web browser

Webcam eye tracking allows researchers to study eye movements in large, geographically diverse online samples, but its accuracy has limited it to coarse comparisons, such as which half or quadrant of the screen a participant is looking at. We introduce SaccadeJS, an open-source webcam eye tracker that runs entirely in the participant's browser. SaccadeJS uses a neural network to transform images of the eyes to features, which a ridge regression maps to screen positions. The network can be rapidly calibrated by computing new regression weights on a sample of fixations at known locations. To train the model, we collected a new dataset of 2,003 participants who fixated targets on their own laptop and desktop computers. In a validation study with 200 new online participants, the median participant's error was about 4% of the screen diagonal. That accuracy is enough to distinguish many more, and smaller, areas of interest than halves and quadrants. We also developed a timing calibration routine to measure each participant's camera lag. After correction, saccade latencies matched distributions from laboratory experiments. We report how different calibration strategies and image characteristics affect accuracy. SaccadeJS could be run in real time on about one third of the participants in the validation study; for others, the library processes frames in the background while the experiment runs. SaccadeJS can be used as standalone software for webcam eye tracking or in conjunction with jsPsych through a set of provided plugins for calibration, validation, and recording.