Measuring live video quality with minimal complexity, now available for everyone!
summary
Claims that ML can predict VMAF and SSIM from encoder internals with high fidelity, using shallow networks and features like motion stats, blur, QP and decoded frames to avoid full-pixel analysis. Implementation targets x264 with an open-source x264 integration and shared code. Reported accuracy is upper-90s correlation and low absolute error versus true VMAF, with runtime about 40× faster than VMAF and only ~2–3% encoder overhead, proven on diverse content.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 3:59Video quality metrics evolution
- 6:49Faster live video metrics
- 8:55Introducing PVMF
- 11:03Conclusion
related talks
33mMeasuring Perceptual Video Quality at Scale
17mContent Aware Encoding for low latency live streaming encoders using deep learning
18mBlind Quality Assessment for Real-Time Visual Communications
26mObjectionable Uses of Objective Quality Metrics
21mWhat to do after per-title encoding
18mCoVA: Exploiting Compressed-Domain Analysis to Accelerate Video Analytics
24mSuper Resolution: The scaler of tomorrow, here today!
10mSubjective video quality assessment for mobile devices