SMAM: Self and Mutual Adaptive Matching for Skeleton-Based Few-Shot Action Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Skeleton-based action recognition is challenging due to sparse data.
• Few-shot learning requires highly discriminative features for similar actions.

Purpose of the Study:

• To develop a novel method for skeleton-based few-shot action recognition.
• To improve the discriminative power of skeleton features for distinguishing similar actions.

Main Methods:

• Propose a self and mutual adaptive matching (SMAM) module.
• Extract features using temporal joint information and spatial relationships.
• Adaptively match features between labeled and query samples.

Main Results:

• SMAM-Net significantly outperforms baseline methods on the NTU RGB + D 120 dataset.
• Achieved state-of-the-art results in one-shot and five-shot action recognition.
• Demonstrated strong generalization across multiple datasets (NTU RGB + D 60, SYSU, PKU-MMD).

Conclusions:

• The proposed SMAM module effectively enhances feature discriminability for few-shot action recognition.
• SMAM-Net offers a reliable and generalizable solution for skeleton-based action recognition tasks.
• Publicly releasing codes and pretrained models to facilitate further research.