Appearance-Based 3D Gaze Estimation with Personal Calibration
share
We propose a way to incorporate personal calibration into a deep learning model for video-based gaze estimation. Using our method, we show that by calibrating six parameters per person, accuracy can be improved by a factor of 2.2 to 2.5. The number of personal parameters, three per eye, is similar to the number predicted by geometrical models. When evaluated on the MPIIGaze dataset, our estimator performs better than person-specific estimators. To improve generalization, we predict gaze rays in 3D (origin and direction of gaze). In existing datasets, the 3D gaze is underdetermined, since all gaze targets are in the same plane as the camera. Experiments on synthetic data suggest it would be possible to learn accurate 3D gaze from only annotated gaze targets, without annotated eye positions.
READ FULL TEXT