Region-of-interest prediction for interactively streaming regions of high resolution video

This paper investigates region-of-interest (ROI) prediction strategies for a client-server system that interactively streams regions of high resolution video. ROI prediction enables pro-active pre-fetching of select slices of encoded video from the server to allow low latency of interaction despite the delay of packets on the network. The client has a buffer of low resolution overview video frames available. We propose and study ROI prediction schemes that can take advantage of the motion information contained in these buffered frames. The system operates in two modes. In the manual mode, the user interacts actively to view select regions in each frame of video. The ROI prediction in this mode aims to reduce the distortion experienced by the viewer in his desired ROI. In the tracking mode, the user simply indicates an object to track and the system supplies an ROI trajectory without further interaction. For this mode, the prediction aims to create a smooth and stable trajectory that satisfies the user's expectation of tracking. While the motion information enables the tracking mode, it also improves the ROI prediction in the manual mode.