Gaze estimation in videoconferencing settings

Abstract

Screen-based communication increasingly replaces face-to-face interactions. Gaze information is important for nonverbal communication. Therefore, we investigate how humans (“receivers”) estimate gaze direction of others (“senders”) in videoconferencing-like settings, a major example of screen-based communication. In two online experiments, receivers estimated gaze targets – which were known to the experimenters – from images of senders. As in real videoconferencing settings, receivers had no information about the geometry of the senders' setup (camera position, etc.), but had to rely on the information available on screen. In Experiment 1, we found that gaze-target estimates were more closely related to the actual target positions in the horizontal than in the vertical direction, a bias toward the sender's head position, and some advantage of presenting the same sender in succession. For Experiment 2, we created a new database of sender images, in which the senders' head position in the image and gaze-target position were systematically varied. Additionally, images of natural scenes were presented whose content served as potential gaze targets. At large, we replicated the findings of Experiment 1, and found only little effect of image content on the estimates. As gaze is an important part of intuitive human-machine interaction, our findings bear relevance beyond videoconferencing. • Gaze estimation in videoconference settings is investigated in two online experiments. • Differences between horizontal and vertical estimates are found. • Some effect of repeating the same sender (person whose gaze is estimated) is observed. • There is little effect of content that is shared between the sender and receiver (person who estimates the gaze). • A database of over 6000 images of 8 senders looking at different gaze targets from different positions is made available.

Topics & Concepts