Predicting the region of interest for dynamic foveated streaming

The nonuniform sampling in human visual system (HVS) is used in a video compression technique called foveation in which the region of interest (ROI) is given a higher bitrate. This technique can significantly reduce the network traffic or provide higher quality with similar bitrate. ROI or fovea region can be predicted using offline algorithms with a considerable prediction error. In real-time video streaming scenario, although fovea region can be detected precisely using an eye-tracker device, accessing to this data is not possible on real-time basis due to the network latency. In this paper, we propose a prediction model which uses streaming client´s gaze locations on a set of frames to predict the fovea region on future frames. With this method we achieved 10× higher prediction accuracy compared to the offline model.