Ruth Rosenholtz¹ (), Niall Williams²; ¹NVIDIA Research, ²University of Maryland, College Park

Moving to online meetings inherently impairs our ability to perceive social cues such as the gaze direction of meeting participants. Much of the previous work on human gaze estimation has examined gaze near the forward direction, i.e. toward the camera. However, in-person group interactions include gaze directions over a much wider range. Here we examine the accuracy of estimating the gaze of human-like avatars looking at each other around a table. Gaze direction in this scene varies in magnitude from 22.5 to 157.5 degrees relative to the camera, and from 0 to 67.5 degrees relative to the avatar body. In the realistic condition, each avatar rotated its torso, head, and eyes according to previous work on human gaze behavior. In the simple turn condition, each avatar moved only its head and eyes, by the same angle. We also varied scene richness; richer scenes may provide stronger cues to depth and body angle, while also adding clutter. Ten participants judged the gaze direction of five randomly chosen MetaHuman avatars in mixed conditions for a total of 300 trials. Results show a consistent bias in perceived gaze direction, as previously reported for smaller viewing angles, but not for realistic torso-head-eye orienting. Uncluttered scenes led to better performance. Gaze uncertainty varied modestly over the wide range of gaze directions, as a function of angle relative to the camera (slope <0.05 degrees/degree of viewing angle), and angle relative to the viewer (slope approximately 0.10 degrees/degree of viewing angle). Over all gaze directions studied, subjects estimated gaze with an average standard deviation of 10.6 degrees, as compared to previous work showing standard deviations of 1-3 degrees for gaze directions within 15 degrees of the direction towards the camera.