Dual-Branch Network with a Subtle Motion Detector for Microaction Recognition in Videos

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Microaction recognition is crucial but challenging due to subtle body part motions.
• Existing methods often rely on high-layer CNN features, which lack detailed motion information.
• Detailed information is vital for recognizing subtle microactions.

Purpose of the Study:

• To enhance microaction recognition by effectively learning mid-layer CNN features.
• To develop a novel dual-branch network for improved microaction classification.
• To introduce a subtle motion detector for refining feature extraction.

Main Methods:

• A dual-branch network utilizing both high-layer and mid-layer CNN features.
• A novel subtle motion detector with discriminative spatial-temporal feature learning, parallel multiplier attention, and activation fusion modules.
• Development of a new microaction video dataset with mixed general and subtle motions.

Main Results:

• The proposed method achieves new state-of-the-art performance on two microaction video datasets.
• The approach demonstrates promising performance on general action video datasets.
• Experimental results validate the effectiveness of leveraging mid-layer CNN features for microaction recognition.

Conclusions:

• The proposed dual-branch network effectively enhances microaction recognition by incorporating mid-layer CNN features.
• The novel subtle motion detector significantly improves the ability to discern fine-grained movements.
• This research advances the field of video-based action recognition, particularly for subtle actions.