Learning joints relation graphs for video action recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current video action recognition methods often overlook the crucial relationships between body joints.
• Existing approaches primarily focus on spatial-temporal features or physical joint dependencies, neglecting relational aspects.

Purpose of the Study:

• To propose a novel joint spatial-temporal reasoning (JSTR) framework for enhanced video action recognition.
• To effectively model and leverage the discriminative relations between body joints for improved performance.

Main Methods:

• Constructing a joints spatial relation graph to capture inter-joint positional information.
• Developing an intra-joint temporal relation graph to model temporal dynamics of individual joints.
• Fusing spatial and temporal reasoning features for comprehensive action recognition.

Main Results:

• The proposed JSTR framework demonstrated strong performance on three real-world video action recognition datasets.
• Experimental results validate the effectiveness of modeling joint relations for action recognition tasks.
• The method achieved good performance across diverse datasets, indicating robustness.

Conclusions:

• Modeling joint relations is critical for advancing video action recognition.
• The JSTR framework offers a promising approach for learning discriminative joint spatial-temporal features.
• This work contributes to more accurate and robust human action understanding in videos.