DMMG: Dual Min-Max Games for Self-Supervised Skeleton-Based Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Skeleton-based action recognition is crucial for human-computer interaction.
• Self-supervised learning offers a promising approach to leverage unlabeled skeleton data.
• Existing methods often struggle with generating diverse and challenging training samples.

Purpose of the Study:

• To propose a novel Dual Min-Max Games (DMMG) based self-supervised method for skeleton action recognition.
• To enhance data augmentation in a contrastive learning framework using adversarial games.
• To improve the learning of discriminative action features from unlabeled skeleton data.

Main Methods:

• Developed a Dual Min-Max Games (DMMG) framework incorporating a viewpoint variation game and an edge perturbation game.
• Employed adversarial paradigms for data augmentation on skeleton sequences and graph-structured body joints.
• Generated hard contrastive pairs by varying viewpoints and perturbing connectivity strength among body joints.

Main Results:

• The viewpoint variation game creates diverse skeleton sequences from multiple viewpoints.
• The edge perturbation game generates varied contrastive samples by altering joint connectivity.
• Achieved superior results on widely-used datasets (NTU-RGB+D, NTU120-RGB+D, PKU-MMD) demonstrating effective action feature representation.

Conclusions:

• The proposed DMMG method effectively generates challenging contrastive pairs for self-supervised learning.
• This approach enables the model to learn representative action features and facilitates transfer to downstream tasks.
• The method achieves state-of-the-art performance in skeleton action recognition using unlabeled data.