Reproduce “Generalizable Video Quality Assessment via Weak-to-Strong Learning” 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
Linhan Cao, Wei Sun, Xiangyang Zhu, Kaiwei Zhang, Jun Jia, Yicong Peng, Dandan Zhu, Guangtao Zhai and 1 more
A team from Shanghai Jiao Tong University and Shanghai Artificial Intelligence Laboratory developed a weak-to-strong (W2S) learning paradigm for Video Quality Assessment, enabling models to achieve robust generalization to out-of-distribution content. The method, which leverages diverse pseudo-labels and iterative training, consistently surpassed state-of-the-art supervised models on challenging benchmarks without relying on human-annotated datasets.
Key insights
- Weak-to-strong generalization is empirically feasible for subjective tasks like VQA, allowing a student model to significantly outperform its teachers, especially on out-of-distribution data.
- Diversifying teacher supervision by ensembling multiple VQA models and incorporating synthetic distortion signals, combined with a ranking-based formulation, substantially enhances VQA model generalization.
- An iterative self-teaching process, guided by sample difficulty and confidence regularization, enables cumulative knowledge transfer, leading to robust and progressively improved performance on challenging unseen video data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.