Thinking of How to Use the Gaze Estimation in the Wild

The eye gaze of humans can directly reflect their attention on some objects, so it becomes one of the most critical factors in improving the user experience of HRI (Human-Robot Interaction) or other intelligent systems. Moreover, empowered by deep learning, gaze estimation methods’ accuracy has been improved and can be used in many applications. However, researchers and engineers still encounter some challenges when using state-of-the-art methods in the wild. One of the challenges is personalization, which means data distribution through people is different. Furthermore, the other challenge is the different devices’ distribution which we will focus on in this paper. The different distribution of images or video will need to be clarified for the model to extract features, while the different devices’ distribution will mislead the methods to project the PoG on the space. So in this paper, we propose a semantic representation for gaze(SRG) representation to alleviate this challenge caused by the different distribution of the devices. Based on the experiment, we got comparable results on different datasets. Moreover, this kind of gaze estimation method can be easily used on various devices.