Eye Gaze Prediction in Videos Using Deep Neural Network

Human eye gaze prediction is a challenging problem that has applications in many fields, such as user interface optimization, AR and VR environments, building assistive technologies, optimizing media storage and transmission, and medical applications. Although many studies proposed methods for predicting eye gaze location in images, very few studies have investigated methods of predicting eye gaze location in videos. In this paper, we propose GazeEngine, a deep neural network for predicting eye gaze location in videos. Our model predicts a single gaze location for every frame in the video. In the heart of GazeEngine is a pretrained object detector that extracts salient features from the video frames. Despite the simplicity of the model, preliminary results show that our model performs well for short distance prediction in Coutrot video dataset. Models such as GazeEngine can be applied to predict gaze location to compress video in a continuous streaming setting, where a client is streaming video from server.