Self-Adaptive Graph With Nonlocal Attention Network for Skeleton-Based Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Graph convolutional networks (GCNs) are used for human body skeleton modeling.
• Existing GCN methods overlook latent joint correlations and non-adjacent temporal relationships.

Purpose of the Study:

• To propose an innovative spatial-temporal model, SAGGAN, to address limitations in current GCNs for action recognition.
• To improve the capture of spatial-temporal features in human skeletal data.

Main Methods:

• Introduced a self-adaptive GCN (SAGCN) module with two dynamic topological graphs.
• Incorporated a global attention network with spatial attention (SA) and temporal attention (TA) modules.
• SAGCN constructs graphs for common data characteristics and unique sample patterns.

Main Results:

• The SAGGAN model captures richer features by considering global connections and temporal dynamics.
• The proposed method overcomes defects of standard graph convolution in action recognition.
• Experiments on NTU-60, NTU-120, and Kinetics datasets show superior performance.

Conclusions:

• SAGGAN demonstrates superior performance in human action recognition tasks.
• The model effectively learns both spatial and temporal features, including non-adjacent frames.
• This approach offers a more comprehensive way to model human skeletal dynamics.