Measuring Perceptual Video Quality at Scale
summary
A concrete metric is available: open‑sourced VMAF model that fuses VIF and DLM plus a simple temporal feature via support‑vector regression to predict perceptual scores trained on Netflix 1080p pristine sources and human ground truth. It is integrated into an encoding pipeline for QC, AB tests and codec comparisons, runs near real time for 1080p with C/C++ optimizations, and has known limits: no 4K training, TV‑centric viewing, handling of film grain and black frames, and withheld training clips.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:13The best video quality possible.
- 2:29Ways to measure quality
- 5:22Design a video quality metric
- 7:22Video Multimethod Assessment Fusion
- 11:13Results
- 12:23Sample Clips
- 13:44Mitigates (some) issues of other metrics
- 16:17Known limitations/questions
- 18:12A metric and a framework. VMAF Development Kit (VDK)
- 24:16Ongoing Improvements: VMAF+
- 25:33The Team
related talks
12mMeasuring live video quality with minimal complexity, now available for everyone!
26mObjectionable Uses of Objective Quality Metrics
18mBlind Quality Assessment for Real-Time Visual Communications
51mVisual Quality Assurance Using Vision-Language Models
30mVideo Quality Assessment of User Generated Contents
21mWhat to do after per-title encoding
17mContent Aware Encoding for low latency live streaming encoders using deep learning
20mStreaming at 250 kbps: raise your expectations