Data-Free Class-Incremental Gesture Recognition With Prototype-Guided Pseudo-Feature Replay.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Existing gesture recognition systems struggle with unseen gestures due to close-set training limitations.
• Class-incremental learning is crucial for adapting gesture recognition to evolving, dynamic environments.

Purpose of the Study:

• To develop a data-free class-incremental learning framework for robust gesture recognition.
• To address the challenge of catastrophic forgetting in continual learning scenarios for gesture recognition.

Main Methods:

• Introduced the Prototype-Guided Pseudo Feature Replay (PGPFR) framework.
• Utilized Pseudo Feature Generation with Batch Prototypes (PFGBP) for dynamic pseudo-feature creation.
• Employed Variational Prototype Replay and Truncated Cross-Entropy for improved learning and stability.
• Implemented Continual Classifier Re-Training to prevent overfitting and maintain feature stability.

Main Results:

• Achieved significant performance gains on SHREC 2017 3D and EgoGesture 3D datasets.
• Demonstrated superior mean global accuracy compared to state-of-the-art methods by 11.8% and 12.8% respectively.
• The PGPFR framework effectively mitigates catastrophic forgetting in incremental gesture learning.

Conclusions:

• The PGPFR framework offers a robust solution for data-free class-incremental gesture recognition.
• This approach enhances the adaptability and longevity of gesture recognition systems.
• The proposed method shows strong potential for real-world applications requiring continuous learning of new gestures.