Efficient Deep Learning-Based Viewport Estimation for 360-Degree Video Streaming

Abstract

While Virtual reality is becoming more popular, 360-degree video transmission over the Internet is challenging due to the video bandwidth.Viewport Adaptive Streaming (VAS) was proposed to reduce the network capacity demand of 360-degree video by transmitting lower quality video for the parts of the video that are not in the current viewport.Understanding how to forecast future user viewing behavior is therefore a crucial VAS concern.This study presents a new deep learning-based method for predicting the typical view for VAS systems.Our proposed solution is termed Head Eye Movement oriented Viewport Estimation based on Deep Learning (HEVEL).Our proposed model seeks to enhance the comprehension of visual attention dynamics by combining information from two modalities.Through rigorous experimental evaluations, we illustrate the efficacy of our approach versus existing models across a range of attention-based tasks.Specifically, viewport prediction performance is proven to outperform four reference methods in terms of precision, RMSE, and MAE.