Merge-and-Split Graph Convolutional Network for Skeleton-Based Interaction Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional graph convolution networks often overlook crucial correlation features between body parts in interaction recognition.
• Capturing these inter-body part dependencies is vital for accurate human interaction understanding.

Purpose of the Study:

• To develop an innovative approach for human interaction recognition that effectively captures correlation features between different body parts.
• To address the limitations of existing methods by treating interaction recognition as a global problem.

Main Methods:

• Introduced the Merge-and-Split Graph Convolutional Network (MS-GCN) that utilizes a novel Merge-and-Split Graph structure.
• Developed a Merge-and-Split Graph Convolution module to extract interaction features by combining the graph structure with Graph Convolutional Networks.
• Incorporated a Short-term Dependence module for joint and motion characteristics and a Hierarchical Guided Attention Module for inter-hierarchical set correlations.

Main Results:

• Achieved state-of-the-art performance on the NTU60 and NTU120 human interaction recognition datasets.
• Demonstrated the model's efficacy in capturing complex correlation features between interaction body parts.
• Validated through extensive experiments, confirming the model's robustness and effectiveness.

Conclusions:

• The proposed MS-GCN model significantly advances human interaction recognition by effectively capturing body part correlations.
• The novel architectural components enable the extraction of rich semantic and temporal information crucial for interaction understanding.
• The open-sourced code facilitates further research and development in the field.