Reproduce “LOGO: A Long-Form Video Dataset for Group Action 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
Shiyi Zhang, Wenxun Dai, Sujia Wang, Xiangwei Shen, Jiwen Lu, Jie Zhou, Yansong Tang
Researchers from Tsinghua University introduce LOGO, the first long-form video dataset for Group Action Quality Assessment (AQA) featuring multi-person artistic swimming performances with detailed action, formation, and score annotations. They also propose GOAT, a Group-aware Attention model, which significantly enhances existing AQA methods by explicitly modeling group interactions and long-term temporal dynamics, achieving improved performance on LOGO and other multi-person datasets.
Key insights
- Explicitly modeling spatial relationships among actors through group features (GCN) is more effective for AQA than relying on simpler spatial configurations (formation features) alone.
- A dedicated temporal fusion strategy, such as self-attention guided by group features, is crucial for effectively processing and aggregating information from long-form videos, outperforming simple averaging.
- The combination of group-aware spatial reasoning and context-sensitive temporal attention provides a generalizable framework that significantly improves existing AQA methods in complex multi-person and long-duration scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.