Saliency-Aware Edge-Based 3D Gaussian Splatting for Immersive Video Streaming

Abstract

3D Gaussian Splatting (3DGS) enables high-quality and efficient novel-view rendering, making it a promising representation for immersive scene streaming. However, supporting interactive 3DGS viewing on lightweight clients over wireless edge networks remains challenging, as view-dependent rendering workload, video encoding, bandwidth fluctuation, and motionto-photon latency are tightly coupled. This paper presents a saliency-aware edge viewport streaming framework for static 3DGS scenes. The proposed framework constructs a reusable 3D saliency prior by accumulating multi-view 2D saliency observations onto Gaussian primitives and sparse voxels. During online streaming, this prior is projected into the current viewport and converted into block-level region-of-interest (ROI) control signals for native video encoding. We further formulate a queueand latency-aware adaptation problem and design a saliencyconditioned recurrent reinforcement learning controller to jointly select the frame rate and target bitrate under dynamic wireless conditions. Experiments on static 3DGS scenes and public 4G LTE traces show that the proposed method improves saliencyweighted viewport quality and Quality of Experience (QoE) over representative baselines, while providing bitrate-efficient native ROI protection. The results also reveal a quality-latency tradeoff compared with more conservative policies.