Skeleton-CutMix: Mixing Up Skeleton With Probabilistic Bone Exchange for Supervised Domain Adaptation

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Domain adaptation in skeleton-based action recognition often relies on complex loss functions for domain alignment.
• Existing methods struggle to leverage the inherent structural properties of skeleton data effectively.
• Mitigating domain shift is crucial for robust performance across different datasets and conditions.

Purpose of the Study:

• To introduce Skeleton-CutMix, a novel augmentation framework for supervised domain adaptation in skeleton-based action recognition.
• To address the limitations of existing domain adaptation techniques by focusing on skeleton data's intrinsic characteristics.
• To improve the generalization and robustness of action recognition models.

Main Methods:

• Developed Skeleton-CutMix, a framework that fabricates mixed-domain skeleton data by exchanging bones between source and target domains.
• Implemented a class-specific bone sampling strategy to prioritize the exchange of important bones for specific action classes.
• Augmented training data with these fabricated skeletons to train more general and robust models.

Main Results:

• Demonstrated the efficiency and effectiveness of the bone exchange augmentation strategy.
• Showcased the preservation of distinctive motion features while successfully mixing action and style across domains.
• Achieved an average gain of over 3% in action recognition accuracy on cross-dataset and cross-age settings (NTU-60, ETRI-Activity3D).

Conclusions:

• Skeleton-CutMix offers a simple yet effective approach to domain adaptation for skeleton-based action recognition.
• The proposed augmentation method outperforms previous domain adaptation techniques and other skeleton augmentation strategies.
• The framework successfully enhances model generalization by leveraging skeleton data's structural correspondences.