Genuine and DeepFake Points of Gaze from a Single Participant During Dyadic Video Calls

This dataset was created to facilitate research on the use of point of gaze tracking for deepfake detection in live video calls. The dataset is composed of the points of gaze for 47 subjects, the facial landmarks of their speaking partner, and speaking context information. For each subject, two deepfakes were generated, one with DeepFaceLive and one with FaceShifter. From each authenticity, the point of gaze was calculated by finding the intersection of the gaze vector, as calculated by MPIIFaeGaze, with the screen. Additionally, the unit vector and magnitude pair from the point of gaze to each of their partners' 478 facial landmarks, as calculated by Google`s Mediapipe, is recorded along with the raw facial landmarks. The speaking context information is also included, which indicates which participant was speaking when this was recorded via a combination of Pyannote and manual annotations.