Reproduce “Towards Unified Video Quality Assessment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chen Feng, Tianhao Peng, Fan Zhang, David Bull
Unified-VQA, developed by the Visual Information Lab at the University of Bristol, provides a single model for both global video quality assessment and diagnostic feedback on specific artifacts across diverse formats like HD, UHD, HDR, and HFR. This framework achieved superior performance against over 18 benchmarks in quality assessment, securing the highest SROCC across all content groups, and demonstrated state-of-the-art artifact detection.
Key insights
- Monolithic VQA models suffer from "representational compromise" when learning features across fundamentally disparate perceptual domains, leading to compromised performance.
- A Diagnostic Mixture-of-Experts (MoE) architecture with semantically distinct experts can effectively disentangle perceptual domains, enabling specialized and robust feature learning for diverse distortion types.
- Multi-task learning, specifically incorporating diagnostic artifact detection as a regularization task, significantly enhances the robustness, disentanglement, and semantic meaningfulness of the VQA backbone's features, thereby improving global quality prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.