Action Shuffling for Weakly Supervised Temporal Localization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised action localization is crucial for video analysis, identifying actions and their temporal spans using only video-level labels.
• Existing methods face challenges due to the order-sensitive and location-insensitive nature of actions.

Purpose of the Study:

• To enhance weakly supervised action localization performance.
• To address the challenges posed by action order sensitivity and location insensitivity.

Main Methods:

• Proposed a novel two-branch network, ActShufNet, incorporating intra-action and inter-action shuffling.
• Intra-action shuffling uses a self-supervised order prediction task to improve video representation.
• Inter-action shuffling augments the training set by reorganizing existing action content.
• Implemented global-local adversarial training to improve robustness against irrelevant noise.

Main Results:

• ActShufNet demonstrated significant efficacy on three benchmark datasets.
• The proposed self-augmented learning framework effectively improved action localization performance.
• The intra/inter-action shuffling and adversarial training contributed to enhanced results.

Conclusions:

• The developed ActShufNet framework successfully improves weakly supervised action localization.
• The novel shuffling strategies and adversarial training are effective for enhancing video analysis tasks.
• The method provides a robust and efficient solution for identifying actions and temporal intervals in videos.