Shallow Graph Convolutional Network for Skeleton-Based Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Graph convolutional networks (GCNs) are effective for skeleton-based action recognition.
• Existing GCN methods use fixed spatial graph sizes, limiting receptive fields and information exploitation.
• Fixed graph sizes lead to redundant action representations and hinder focus on beneficial features.

Purpose of the Study:

• To address limitations of fixed graph sizes in GCNs for action recognition.
• To propose an adaptive module for efficient merging of skeleton graph vertices.
• To develop a novel, computationally efficient GCN model for improved action recognition.

Main Methods:

• Introduced a plug-and-play channel adaptive merging module (CAMM) for human skeleton graphs.
• CAMM adaptively and efficiently merges vertices from the same skeleton parts with channel-specific weights.
• Developed a shallow graph convolutional network (SGCN) utilizing the CAMM.

Main Results:

• The proposed CAMM allows flexible joint integration across different channels.
• The SGCN model achieves state-of-the-art performance on action recognition tasks.
• The method demonstrates superior results with reduced computational cost on benchmark datasets.

Conclusions:

• The CAMM effectively overcomes the limitations of fixed graph sizes in GCNs.
• The SGCN model offers an efficient and high-performing solution for skeleton-based action recognition.
• The proposed approach enhances the model's ability to exploit discriminative information and reduces redundancy.