Analytic Score Optimization for Multi Dimension Video Quality Assessment
New method uses GPT-generated rationales and a novel optimization objective to beat closed-source APIs on video quality prediction.
Deep Dive
Researchers from multiple institutions introduced UltraVQA, a large-scale dataset with videos annotated across 5 quality dimensions (Motion, Aesthetic, Content, etc.) by over 3 human raters each, plus GPT-generated rationales. They also developed Analytic Score Optimization (ASO), a post-training objective with a closed-form solution that captures ordinal human ratings. The method outperforms most baselines, including closed-source APIs, and reduces mean absolute error in quality prediction.
Why It Matters
Enables more nuanced, interpretable AI evaluation of video content for platforms, creators, and streaming services.