YuZu: Neural-Enhanced Volumetric Video Streaming
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
System applies optimized 3D super-resolution to point-cloud volumetric video, achieving 542× speedup on SR inference through patch-generation trimming, model pruning, input/output merging, and patch-level caching via dynamic programming. Prototype streams at 30+ FPS, cuts GPU memory by ~87%, and uses a QoE model validated in a 1,446-subject study with median prediction error 12.49%. End-to-end evaluations show 37–178% QoE gain over no SR and 101–175% improvement versus viewport-adaptive baselines.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:24Volumetric Video
- 4:42Our Approach: Yuzu
- 5:18YuZu Overview Server
- 7:41Intra-frame SR
- 12:07SR Performance Breakdown
- 13:10Yuzu's QoE Improvement
- 14:17YuZu's End-to-end Performance
- 15:16YuZu vs. Viewport-Adaptive Streaming
- 16:05Demo
related talks
21mReal-Time Video Super Resolution for Live Streaming with WebGPU
24mSuper Resolution: The scaler of tomorrow, here today!
17mGemino: Practical and Robust Neural Compression for Video Conferencing
33mArchitecting Scalable Multi-Modal Video Search
17mMorphe: High-Fidelity Generative Video Streaming with Vision Foundation Model
18mVideo @Scale Fall 2022 LIVE Q&A #3
35mLarge-Scale Video Intelligence
18mGaze-Driven Video Delivery: Science Fiction or Viable Link to the Metaverse?