TDP-DETR: Temporal dynamics perception framework for video moment retrieval and highlight detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Video Moment Retrieval (VMR) and Highlight Detection (HD) are crucial for analyzing untrimmed videos.
• Accurate temporal boundary perception is essential for VMR and HD tasks.
• Current models struggle with precise action semantic alignment and exhibit boundary perception bias due to dynamic visual cues.

Purpose of the Study:

• To propose a novel Temporal Dynamics Perception DEtection TRansformer (TDP-DETR) for improved VMR and HD.
• To address limitations in current models regarding temporal dynamics and boundary perception.
• To enhance action semantic alignment and reduce boundary bias in video analysis.

Main Methods:

• TDP-DETR models action temporal dynamics from temporal persistence and temporal progression perspectives.
• A dynamic masking strategy is used for action duration-aware temporal modeling, incorporating temporal priors.
• An action state difference perception module captures frame-to-frame variations to perceive action progression speed.

Main Results:

• The proposed TDP-DETR method consistently outperforms existing state-of-the-art approaches on three MR/HD benchmarks.
• The approach demonstrates improved temporal boundary perception and action semantic alignment.
• The dynamic masking and action state difference modules contribute to enhanced performance.

Conclusions:

• TDP-DETR effectively models action temporal dynamics for precise boundary prediction in VMR and HD.
• The method offers a significant advancement in handling temporally dynamic video content.
• Future work will involve making the code publicly available for further research.