DeGCN: Deformable Graph Convolutional Networks for Skeleton-Based Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Skeleton-based action recognition leverages human body graph topology.
• Existing Graph Convolutional Networks (GCNs) use inflexible message aggregation, failing to account for intra-class variations and redundant connections in skeleton sequences.

Purpose of the Study:

• To propose a novel Deformable Graph Convolutional Network (DeGCN) for adaptive and accurate skeleton-based action recognition.
• To address limitations of current GCNs in handling diverse action samples and noisy skeleton data.

Main Methods:

• Developed Deformable Graph Convolutional Network (DeGCN) with deformable sampling on spatial and temporal graphs.
• Incorporated continuous latent space for temporal features to capture inherent action continuity.
• Designed an innovative multi-branch framework for joint and bone modalities, balancing accuracy and model size.

Main Results:

• Achieved state-of-the-art performance on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.
• Demonstrated DeGCN's ability to adaptively capture informative joints and perceive discriminative receptive fields.
• Showcased the effectiveness of the multi-branch framework in enhancing ensemble effects and model efficiency.

Conclusions:

• The proposed DeGCN method significantly advances skeleton-based action recognition.
• Deformable graph convolution offers a more suitable approach for analyzing complex human actions from skeleton data.
• The multi-branch framework provides a robust and efficient solution for action recognition tasks.