Contextual Bandit Learning-Based Viewport Prediction for 360 Video

Abstract

Accurately predicting where the user of a Virtual Reality (VR) application will be looking at in the near future improves the perceive quality of services, such as adaptive tile-based streaming or personalized online training. However, because of the unpredictability and dissimilarity of user behavior it is still a big challenge. In this work, we propose to use reinforcement learning, in particular contextual bandits, to solve this problem. The proposed solution tackles the prediction in two stages: (1) detection of movement; (2) prediction of direction. In order to prove its potential for VR services, the method was deployed on an adaptive tile-based VR streaming testbed, for benchmarking against a 3D trajectory extrapolation approach. Our results showed a significant improvement in terms of prediction error compared to the benchmark. This reduced prediction error also resulted in an enhancement on the perceived video quality.

Tile-Based Edge Caching for 360° Live Video Streaming

“…To maximize the quality of the 360 o videos displayed by the users, authors in [22] optimize the video encoding configurations so that 360 o videos are optimally encoded into multiple representations. To predict the demanded viewports by the users in the near future, authors in [23] present a contextual bandit algorithm. Differently from [23], authors in [24] propose a trajectory-based viewport prediction algorithm, aiming to predict the users' requested viewports in the long-term.…”

“…To predict the demanded viewports by the users in the near future, authors in [23] present a contextual bandit algorithm. Differently from [23], authors in [24] propose a trajectory-based viewport prediction algorithm, aiming to predict the users' requested viewports in the long-term. To deal with the problem of inaccurate distortion measurements occurring when 360 o videos are projected from the spherical domain to 2D plane, the work in [25] proposes a method that optimizes the encoding process based on signals' distortion in spherical domain.…”

SVP: Sinusoidal Viewport Prediction for 360-Degree Video Streaming

“…However, all these methods are not accurate in the yaw direction due to the periodicity issue. Heyse et al [8] propose a contextual bandit based approach to estimate user's future viewport. The accuracy of this method can be lower than those regression-based methods, since it does not make use of historical information.…”

Human-centric Quality Management of Immersive Multimedia Applications

“…Type Method Heyse et al [24] Content-agnostic Contextual Bandits Petrangeli et al [25] Content-agnostic Trajectory-based + clustering Van der Hooft et al [26] Content-agnostic Further refinement of [25] Ban et al [9] Content-agnostic LR + probability voting Fan et al [27] Hybrid Saliency analysis & motion detection + CNNs/LSTMs on historical viewport orientations Jeong et al [28] Content-agnostic 360°audio location as predictor for next viewport diction. Content-based prediction aims at characterizing the given content, independent from the user, in terms of saliency maps, motion detection, Regions-Of-Interest (ROIs) etc.…”

“…static prediction (the user's orientation does not change between two consecutive samples) or 3D linear interpolation (the user's movement between the current and the next sample is the same as between the previous and the current) to more complex ML-based methods based on ANNs or SVMs [23]. An overview of these approaches is provided in Table III. Heyse et al [24], for example, present a content-agnostic Reinforcement Learning (RL) approach based on Contextual Bandits (CBs). To this extent, a set of previous HMD orientations is fed to two learners per spatial dimension (polar angle and azimuth).…”