Improving the Reliability of Gaze Estimation through Cross-dataset Multi-task Learning
Hanlin Zhang, Xinming Wang, Weihong Ren, Bernd Rainer Noack, Honghai Liu · 2022
In recent years, gaze estimation has been applied to numerous application scenarios, such as driver monitor systems and autism evaluation, to name a few. However, the current algorithms are based on an assumption: the users eyes are all open in the input images. In practice, this assumption does not always hold since people blink. Images with eyes closed will result in irregular gaze estimation results that will affect the safety of the vehicle driving and the reliability of the evaluation results. Here, we address the problem by leveraging cross-dataset training to train a multi-task network capable of simultaneously estimating the eye gaze and the blink on the gaze estimation dataset and blink estimation dataset. We further compare our networks results with that of single-task networks and evaluate our network in a simulated scenario. The experimental results demonstrate that our method is suitable for practical use and can improve the reliability of the gaze estimation results.