Reproduce “Learning Group Activity Features Through Person Attribute Prediction” on OpenResearch

About this paper

Chihiro Nakatani, Hiroaki Kawashima, Norimichi Ukita

Researchers from Toyota Technological Institute and University of Hyogo developed a self-supervised framework for learning Group Activity Features (GAFs) by predicting individual person attributes. This approach achieved state-of-the-art performance in group activity retrieval and demonstrated the ability to learn fine-grained distinctions within activity classes, significantly reducing reliance on complex group activity annotations.

Key insights

• Supervising GAF learning with individual person attribute prediction, particularly when guided by spatial location, provides an effective signal for capturing complex group dynamics.
• Masking individual features during training enhances the GAF's ability to learn richer inter-person interactions and robust contextual representations.
• The learned GAF is capable of representing fine-grained distinctions, allowing for the implicit discovery of meaningful sub-categories within broader group activities without explicit supervision.

Summary by alphaXiv.