Action Recognition With Motion Diversification and Dynamic Selection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Action recognition relies heavily on motion modeling.
• Varying motion dynamics (tempo, amplitude) in videos challenge current methods.
• Adaptive motion information capture is a key research area.

Purpose of the Study:

• To develop a novel module for adaptive motion feature extraction in action recognition.
• To address the challenge of diverse motion dynamics in video clips.
• To improve the accuracy and efficiency of action recognition systems.

Main Methods:

• Introduced a Motion Diversification and Selection (MoDS) module.
• Proposed a spatio-temporal motion generation (StMG) module to create diverse motion features.
• Utilized a dynamic motion selection (DMS) module to identify the most discriminative features.

Main Results:

• The MoDS module effectively generates and selects spatio-temporal motion features.
• Achieved state-of-the-art performance on Something-Something V1 & V2 benchmarks.
• Demonstrated effectiveness across five widely-used action recognition benchmarks.

Conclusions:

• The proposed MoDS module enhances action recognition by leveraging diversified motion information.
• The method maintains computational efficiency during inference.
• This approach is particularly effective for videos with significant motion variations.