HSMF-Net: Semantic Viewport Prediction for Immersive Telepresence and On-Demand 360-degree Video
The acceptance of immersive telepresence systems is impeded by the latency that is present when mediating the realistic feeling of presence in a remote environment to a local human user. A disagreement between the user's ego-motion and the visual response provokes the emergence of motion sickness. Viewport or head motion (HM) prediction techniques play a key role in compensating the noticeable delay between the user and the remote site. We present a deep learning-based viewport prediction paradigm that fuses past HM trajectories with scene semantics in a late-fusion manner. Real HM profiles are used to evaluate the proposed approach. A mean compensation rate as high as 99.99% is obtained, clearly outperforming the state-of-the-art. An on-demand 360-degree video streaming framework is presented to prove its general validity. The proposed approach increases the perceived video quality while requiring a significantly lower transmission rate.