Learning Semantics-Preserving Attention and Contextual Interaction for Group Activity Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional group activity recognition methods often use simple pooling, losing crucial contextual and semantic information.
• Existing approaches struggle to effectively identify key individuals and their interactions within a group.

Purpose of the Study:

• To develop an advanced architecture for group activity recognition that preserves semantic information and models contextual interactions.
• To improve the accuracy and efficiency of recognizing group activities from video data.

Main Methods:

• Introduced a Semantics-Preserving Teacher-Student (SPTS) network architecture.
• Employed a Teacher Network in the semantic domain and a Student Network in the appearance domain, enforcing the latter to mimic the former.
• Integrated graph convolutional modules to model inter-person dependencies within both networks.
• Extended the approach for action segmentation using intermediate features.

Main Results:

• The SPTS network effectively allocates semantics-preserving attention, identifying key individuals and discarding misleading ones.
• Graph convolutional modules successfully reason inter-person dependencies.
• Experimental results on four datasets demonstrate superior performance compared to state-of-the-art methods.
• The approach shows effectiveness in action segmentation tasks.

Conclusions:

• The proposed SPTS network offers a significant advancement in group activity recognition by effectively leveraging semantic preservation and contextual modeling.
• The method achieves state-of-the-art performance without requiring additional labeled data.
• The integration of graph convolutional networks enhances the understanding of complex group dynamics.