Active Factor Graph Network for Group Activity Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Group activity recognition is crucial for understanding complex human behaviors.
• Existing methods often rely on limited second-order interactions between individuals.
• This limitation hinders accurate recognition of nuanced group dynamics.

Purpose of the Study:

• To propose a novel method for group activity recognition that captures higher-order interactions.
• To address the insufficiency of second-order interaction modeling in current approaches.
• To enhance the accuracy and interpretability of group activity recognition systems.

Main Methods:

• Introduced a third-order active factor graph network to model interactions among three individuals.
• Developed an active individual selection mechanism based on influence weights to reduce noise.
• Designed a two-branch network (full and active factor graphs) and a consistency-aware reasoning module.

Main Results:

• Achieved state-of-the-art performance on four benchmark datasets: Volleyball, Collective Activity, Collective Activity Extended, and SoccerNet-v3.
• Demonstrated the effectiveness of modeling third-order interactions for improved group activity recognition.
• Visualization results confirmed the interpretability of the proposed method.

Conclusions:

• The proposed third-order active factor graph network effectively models complex group interactions.
• The method surpasses existing approaches in group activity recognition accuracy.
• The approach offers enhanced interpretability, providing insights into individual contributions to group activities.