Boundary Conditions for Human Gaze Estimation on A Social Robot using State-of-the-Art Models

摘要

Appearance-based methods are a promising solution for gaze estimation, as they eliminate the need for additional devices and calibration. This makes them particularly well-suited for human-robot interaction (HRI) research. However, until recently their performance was under par compared to traditional eye-trackers. Recent breakthroughs have been made with the release of two large-scale datasets with a wide range of gaze directions (Gaze360 and ETH-XGaze) and the accompanying state-of-the-art deep neural networks (L2CS and ETH). In this paper, we systematically evaluate the performance of these two appearance-based models on a social robot. In our setup, we vary the distance from the robot (1-3m) and camera resolution (640 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">*</sup> 480 and 3840 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">*</sup> 2160) and analyze the performance in terms of accuracy and precision. We find that the L2CS model trained on the Gaze360 dataset combined with a 4K camera achieves the best performance on the 2 m and 3 m distances. We show that a simple offset correction on pitch and yaw can further increase the accuracy and precision by 18.6% and 9.6% respectively. We conclude that for a range up to 3 m appearance-based gaze estimation models provide a promising approach for application in HRI research.