Learning Multi-View Interactional Skeleton Graph for Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Skeleton-based action recognition is crucial for understanding human activities.
• Existing graph-based methods struggle with modeling spatial context due to fixed interaction patterns and inflexible GCN weights.
• There is a need for methods that can capture multi-level spatial information in a unified manner.

Purpose of the Study:

• To propose the multi-view interactional graph network (MV-IGNet) for improved skeleton-based action recognition.
• To address the limitations of existing methods in modeling spatial context.
• To develop a unified framework for constructing, learning, and inferring multi-level spatial skeleton context.

Main Methods:

• MV-IGNet constructs multi-level spatial skeleton context (view-level, group-level, joint-level) in a unified way.
• It leverages different skeleton topologies as multi-views to generate complementary action features.
• Separable parametric graph convolution (SPG-Conv) enriches local interaction patterns and adapts to irregular topologies, while a global context adaption (GCA) module learns input-dependent topologies.

Main Results:

• MV-IGNet achieves impressive performance on large-scale benchmarks, including NTU-RGB+D and NTU-RGB+D 120.
• The proposed method demonstrates superior spatial context modeling capabilities compared to mainstream works.
• MV-IGNet offers a smaller model size and faster inference speeds.

Conclusions:

• MV-IGNet effectively captures multi-level spatial context for skeleton-based action recognition.
• The method provides a unified and adaptable framework for handling diverse skeleton topologies.
• MV-IGNet represents a significant advancement in efficient and high-performance action recognition.