PRECEPT: Power-Efficient Field-of-View Prediction for VR Video Streaming
Field-of-view (FoV) prediction is critical for reducing device energy consumption and enhancing user quality of experience (QoE) in immersive streaming. To address the high computational and energy costs of standard DL-based FoV prediction, we propose PRECEPT, an energy-efficient, system-oriented two-stage framework. PRECEPT splits the prediction pipeline by adding a lightweight, CPU-based classifier to identify tile change. PRECEPT's classifier successfully filters approximately 80% of “no-change” events. PRECEPT activates the resource-intensive DL model only during identified tile change. This design reduces the average inference delay and energy consumption by up to 69% in a real mobile deployment. PRE-CEPT's two-stage design enables sustainable, high-performance FoV prediction on resource-constrained devices.