Spatial-temporal graph neural ODE networks for skeleton-based action recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Skeleton-based action recognition is vital for virtual reality and motion analysis.
• Challenges include intraindividual action variations and long-term temporal dependencies.
• Existing models struggle to effectively address these complexities.

Purpose of the Study:

• To propose an innovative model, spatial-temporal graph neural ordinary differential equations (STG-NODE), for enhanced skeleton-based action recognition.
• To address challenges of intraindividual action differences and long-term temporal dependencies.
• To improve the accuracy and temporal modeling capabilities of action recognition systems.

Main Methods:

• Utilized dynamic time warping (DTW) for normalizing 3D skeleton data and deriving custom adjacency matrices.
• Applied a custom ordinary differential equation (ODE) integrator to simulate the dynamic evolution of temporal features.
• Employed an ODE solver to numerically process time features, enhancing long-term dependency modeling.

Main Results:

• STG-NODE demonstrated superior performance on the NTU RGB+D 60 and Kinetics Skeleton 400 benchmark datasets.
• The model effectively improved recognition accuracy by addressing intraindividual differences and temporal dependencies.
• Achieved more powerful temporal modeling capabilities compared to existing methods.

Conclusions:

• The STG-NODE model offers a significant advancement in skeleton-based action recognition.
• The approach provides effective solutions for handling complex action variations and temporal dynamics.
• Presents new methodologies for future research and development in the action recognition field.