Host-Parasite: Graph LSTM-in-LSTM for Group Activity Recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Group activity recognition in multi-person scenes is challenging.
• Existing Long Short-Term Memory (LSTM)-based methods use a two-stage approach, neglecting spatiotemporal relationships.
• This neglects the inherent
• host-parasite
• dynamic between group activities and individual actions.

Purpose of the Study:

• To propose a novel Graph LSTM-in-LSTM (GLIL) model for group activity recognition.
• To simultaneously model person-level actions and group-level activity.
• To address the limitations of existing two-stage recognition methods.

Main Methods:

• Introduced a novel Graph LSTM-in-LSTM (GLIL) architecture.
• Modeled group activity using a
• host-parasite
• framework with simultaneous person-level (P-LSTMs) and group-level (G-LSTM) modeling.
• Incorporated a residual LSTM with residual connections to learn person-level temporal and static features.

Main Results:

• The proposed GLIL model effectively recognizes group activities in multiple-person scenes.
• GLIL demonstrated superior performance compared to state-of-the-art methods on two public datasets.
• The
• host-parasite
• architecture and simultaneous modeling proved effective.

Conclusions:

• The GLIL model offers a significant advancement in group activity recognition.
• Simultaneous modeling of person-level and group-level activities is crucial for capturing spatiotemporal dynamics.
• GLIL provides a more effective approach to understanding complex human interactions in group settings.