A Comprehensive Evaluation of OpenFace 2.0 Gaze Tracking.

Gaze tracking is widely used for various human-computer interaction (HCI) applications. One of the exciting pieces of gaze tracking and analysis software is Openface 2.0, an open-source and powerful facial landmark tracking toolkit that enables real-time head pose tracking, eye gaze estimation, and action unit recognition with webcams. However, despite its various advantages, many researchers are concerned about the low accuracy and unstable performance of OpenFace 2.0, mainly gaze tracking. Indeed, the authors of OpenFace describe their gaze tracking as an estimation instead of accurate computation, with certain limitations which are not systematically explored or explained by the authors or previous research. Therefore, this paper aims to evaluate OpenFace 2.0 gaze tracking under various experimental settings. Specifically, the results may provide insightful information about how OpenFace 2.0 gaze-tracking performance may change when conditions such as the distance between users and the camera, lighting, camera position, user head pose, and facial obfuscation vary. The evaluation could especially benefit researchers who intend to use OpenFace 2.0 gaze tracking in less favorable environments and settings.