Device-Adaptive 2D Gaze Estimation: A Multi-Point Differential Framework.

Eye tracking system on mobile devices is important for many interactive applications. However, since models are usually customized with limited types of devices and new devices have totally different physical parameters, it is hard to generalize over unseen devices. In this paper, we present a device-adaptive 2D gaze estimation algorithm based on differential prediction. We reformulate the gaze estimation as a relative position prediction problem between the input image and calibration images, which skips the estimation for camera parameters and makes models easily generalize over devices. To tackle the new challenge, this work proposes a framework which jointly trains a differential prediction module and an aggregation module for ensembling the predictions from multiple calibration points. Experiments show that the framework outperforms baseline models constantly on open datasets with only 3–5 calibration points.