Motion Attention-Guided Relational Reasoning for Weakly Supervised Group Activity Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing attention-based methods for weakly supervised group activity recognition struggle with generating diverse token embeddings.
• Label-free approaches require robust mechanisms to identify and focus on relevant actors within a scene.

Purpose of the Study:

• To develop a novel framework for weakly supervised group activity recognition that enhances token embedding diversity.
• To improve the focus on actor motion regions for more accurate activity recognition.

Main Methods:

• A Motion-Guided Mask Generator (MGMG) module was developed to estimate attention region masks using grayscale motion masks derived from motion direction.
• MGMG incorporates a correlation layer, cosine attention, a mask generator, and a specialized activation function.
• A normalized relative error loss function was customized for the MGMG module to handle value range mismatches.
• A Motion Attention-Guided Relational Reasoning (MAGRR) framework was proposed, utilizing MGMG and a Spatial-temporal Aggregation Stack (SAS) for attention region activation and temporal dependency capture.

Main Results:

• The proposed MAGRR framework achieved state-of-the-art performance on the Collective Activity and Collective Activity Extension datasets.
• Competitive performance was demonstrated on the Volleyball and NBA datasets.
• The MGMG module effectively generated diverse token embeddings and focused attention on actor motion.

Conclusions:

• The developed MGMG and MAGRR framework significantly advances weakly supervised group activity recognition.
• The approach provides a robust solution for attention-based methods facing challenges with token embedding diversity and actor focus.