Blind Quality Assessment for Real-Time Visual Communications
summary
A lightweight ML pipeline for blind real‑time video quality assessment runs on end devices and claims near state‑of‑the‑art correlation with human mean opinion scores by trading a few percent of accuracy for huge efficiency gains. It extracts DCT‑based spatial features, optional 3D spatio‑temporal cubes for motion, applies a partitioned MSE feature selection and a regressor, yielding GreenBIQA/GreenBBQA models that are about 50x smaller and ~500–700x less compute than top deep nets, but still mispredict attention‑focused edge cases.
AI-written summary · about
chapters (7)
- 0:00Blind video quality assessment intro
- 1:32Netflix and Meta collaboration
- 4:04Blind image quality methods
- 9:32Image model performance comparison
- 13:27Extending to blind video evaluation
- 15:54Video model performance comparison
- 16:39Conclusion and future challenges
related talks
12mMeasuring live video quality with minimal complexity, now available for everyone!
51mVisual Quality Assurance Using Vision-Language Models
33mMeasuring Perceptual Video Quality at Scale
34mTesting Live Video Streams At Scale
36mAI-Powered Real-Time Quality Control
35mLarge-Scale Video Intelligence
41mPractical Aspects of Applying Modern Computer Vision Technologies
17mContent Aware Encoding for low latency live streaming encoders using deep learning