Dual-Recommendation Disentanglement Network for View Fuzz in Action Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-view action recognition identifies actions from multiple perspectives.
• Existing methods struggle with fuzzy view-action relationships, leading to recognition errors.
• Disentangling view and action components is crucial for accurate recognition.

Purpose of the Study:

• To propose a novel Dual-Recommendation Disentanglement Network (DRDN) for enhanced multi-view action recognition.
• To mitigate the negative impact of fuzzy views on action recognition accuracy.
• To develop a method that effectively disentangles view and action features.

Main Methods:

• The Dual-Recommendation Disentanglement Network (DRDN) models images as compositions of view and action.
• Specific Information Recommendation (SIR) enhances action representation by considering intricate activities and views.
• Pyramid Dynamic Recommendation (PDR) learns global view representations by transferring features and resisting noise.

Main Results:

• DRDN achieves state-of-the-art performance on multiple standard benchmarks.
• The proposed method demonstrates superior accuracy in multi-view action recognition compared to existing approaches.
• Experiments validate the effectiveness of SIR and PDR in disentangling view and action features.

Conclusions:

• The DRDN effectively addresses the challenge of fuzzy views in multi-view action recognition.
• The proposed SIR and PDR modules contribute to complete action and view representations.
• DRDN offers a robust and accurate solution for multi-view action recognition tasks.