Micro-DualNet: Dual-Path Spatio-Temporal Network for Micro-Action Recognition

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Biomedical Engineering

Background:

• Micro-actions are brief, subtle movements crucial for social communication and natural human interactions.
• Current computer vision systems lack understanding of these fine-grained actions due to their diverse spatio-temporal characteristics.
• Existing methods struggle to capture both spatial configurations and temporal dynamics inherent in micro-actions.

Purpose of the Study:

• To address the challenge of diverse spatio-temporal characteristics in micro-actions for improved fine-grained video understanding.
• To develop a novel deep learning architecture capable of processing micro-actions with varying spatial and temporal definitions.
• To enhance the performance of computer vision systems in recognizing and interpreting subtle human movements.

Main Methods:

• Proposed a dual-path network with parallel Spatial-Temporal (ST) and Temporal-Spatial (TS) pathways.
• Implemented anatomically-grounded spatial entity processing.
• Introduced entity-level adaptive routing and Mutual Action Consistency (MAC) loss for cross-path coherence.

Main Results:

• Demonstrated competitive performance on the MA-52 dataset.
• Achieved state-of-the-art results on the iMiGUE dataset.
• Validated the effectiveness of the dual-path architecture and adaptive routing for micro-action recognition.

Conclusions:

• Architectural adaptation is essential for handling the complexity of micro-actions in fine-grained video understanding.
• The proposed dual-path network effectively captures diverse spatio-temporal features of micro-actions.
• This research advances the capabilities of computer vision in interpreting subtle human behaviors.