FAU-Gaze: Fast and Accurate User-specific Gaze Estimation Framework

Gaze estimation has a wide range of applications such as neuroscience and clinical research. In this paper, we propose and implement a fast and accurate user-specific gaze estimation system, called FAU-Gaze. FAU-Gaze supports online real-time training with an inference speed of up to 7-11.5 ms in 100 FPS. Compared with existing models, the kernel model FPGC (Feature-based Personalized Gaze Calibrator) of FAU-Gaze increases the accuracy by 36.4% and 33.7% on MPIIFaceGaze and TabletGaze respectively. By mining each user's potential characteristics, FAU-Gaze can more accurately locate each user's real gaze position. In order to test FAU-Gaze, we also introduce a low-resolution and low-definition laptop gaze estimation dataset TobiiGaze containing 41,000 images. Through our experiments on both TobiiGaze, MPIIFaceGaze, and TabletGaze, the prediction error of FAU-Gaze is reduced to 1.61 cm and the robustness outperforms the state-of-the-art.