PARKS-Gaze - A Precision-focused Gaze Estimation Dataset in the Wild under Extreme Head Poses
L. R. D. Murthy, Abhishek Mukhopadhyay, Ketan Anand, Shambhavi Aggarwal, Pradipta Biswas · 2022
The performance of appearance-based gaze estimation systems that utilizes machine learning depends on training datasets. Most of the existing gaze estimation datasets were recorded in laboratory conditions. The datasets recorded in the wild conditions display limited head pose and intra-person variation. We proposed PARKS-Gaze, a gaze estimation dataset with 570 minutes of video data from 18 participants. We captured head pose range of ± 50, [-40,60] degrees in yaw and pitch directions respectively. We captured multiple images for a single Point of Gaze (PoG) enabling to carry out precision analysis of gaze estimation models. Our cross-dataset experiments revealed that the model trained on proposed dataset obtained lower mean test errors than existing datasets, indicating its utility for developing real-world interactive gaze controlled applications.