Video Quality Assessment of User Generated Contents
summary
YouTube releases YT-UGC, ~1,500 raw uploads with MOS/DMOS and 600+ content labels, and trains a deep model called YouVQ that claims higher correlation to human scores than existing no-reference metrics. YouVQ uses an intermediate self-supervised retraining step on synthetic distortions, three sub-networks for content, distortion, and compression, patch-based native-resolution inputs, outputs a single score plus localized artifact heatmaps, and reports distortion-class accuracy ~0.97 and improved generalization across other UGC datasets.
AI-written summary · about
chapters (17)
- 0:00UGC and video quality intro
- 3:22Traditional metrics challenges
- 4:27YouTube UGC dataset
- 7:37Introducing UVQ metric
- 10:06Transfer learning and retraining
- 12:29UVQ model design
- 13:56UVQ performance
- 16:50UVQ quality report example
- 17:53Summary and future
- 18:15Q&A: Open dataset
- 19:40Q&A: Live stream methods
- 21:03Q&A: Retraining frequency
- 21:45Q&A: ROI-based scoring
- 23:33Q&A: UVQ public availability
- 23:54Q&A: Variable frame rates
- 24:37Q&A: AI vs traditional metrics
- 27:05Q&A: Future of quality measurement
related talks
33mMeasuring Perceptual Video Quality at Scale
12mMeasuring live video quality with minimal complexity, now available for everyone!
15mML Algorithms for Trust and Safety
18mBlind Quality Assessment for Real-Time Visual Communications
51mVisual Quality Assurance Using Vision-Language Models
26mObjectionable Uses of Objective Quality Metrics
10mNon-Linear Consumption of Videos Using a Sequence of Personalized Multimodal Fragments
29mBuilding Generative Content Understanding Systems