Related Experiment Video
Updated: May 8, 2025
13:00
Measuring Attention and Visual Processing Speed by Model-based Analysis of Temporal-order Judgments
Published on: January 23, 2017
9.8K
Temporal Gap-Aware Attention Model for Temporal Action Proposal Generation
Sorn Sooksatra1, Sitapa Watcharapinchai1
1National Electronic and Computer Technology Center, National Science and Technology Development Agency, Khlong Nueng, Khlong Luang District, Pathum Thani 12120, Thailand.
Journal of Imaging
|December 27, 2024
Summary
This study introduces an attention mechanism to improve temporal action proposal generation in videos. The method enhances the segmentation of contiguous actions, achieving higher accuracy in boundary localization.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Temporal action proposal generation aims to extract action instances from untrimmed videos.
- Current methods face challenges in accurately segmenting contiguous action proposals with minimal temporal gaps.
Purpose of the Study:
- To enhance the segmentation of contiguous action proposals in untrimmed videos.
- To improve the localization accuracy of action boundaries by addressing temporal gaps.
Main Methods:
- Incorporation of an attention mechanism to weigh proposal importance within contiguous groups.
- Utilizing gap displacement between proposals to compute attention scores for boundary localization.
Main Results:
- Significant improvement in the performance of short-duration and contiguous action proposals.
- Achieved an average recall of 78.22% on ActivityNet v1.3 and Thumos 2014 datasets.
- Outperformed a state-of-the-art boundary-based baseline.
Conclusions:
- The proposed attention mechanism effectively addresses limitations in segmenting contiguous action proposals.
- This approach offers a more accurate method for temporal action localization in videos.
