GRA: Graph Representation Alignment for Semi-Supervised Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Graph convolutional networks (GCNs) are effective for human action recognition using skeletal data.
• Current GCN methods require large labeled datasets, which are costly and difficult to obtain.
• Incomplete skeleton data (missing joints/frames) poses a significant challenge during testing.

Purpose of the Study:

• To develop a novel approach, Graph Representation Alignment (GRA), to address data limitations in GCN-based action recognition.
• To reduce the dependency on extensive labeled datasets for training GCN models.
• To enhance the robustness of GCNs against incomplete skeletal data.

Main Methods:

• Introduced a self-training (ST) paradigm to generate high-quality pseudo-labels, minimizing the need for manual labeling.
• Implemented a representation alignment (RA) technique using consistency regularization to mitigate the impact of missing data.
• Evaluated the GRA approach on NTU RGB+D and N-UCLA benchmarks.

Main Results:

• GRA significantly reduces the requirement for labeled data, enabling stable model training with minimal supervision.
• The representation alignment technique effectively handles incomplete skeleton data, maintaining high performance.
• GRA demonstrated improved GCN performance in data-constrained scenarios and robustness against missing data components.

Conclusions:

• GRA offers a viable solution for data-efficient and robust human action recognition using GCNs.
• The proposed method alleviates the practical challenges associated with large-scale data acquisition and data incompleteness.
• GRA advances the applicability of GCNs in real-world action recognition tasks.