Weakly supervised temporal action localization with actionness-guided false positive suppression.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised temporal action localization (WS-TAL) uses video-level labels for action boundary detection.
• Existing 'localization-by-classification' methods generate false positives due to class-related scene interference.
• Treating background as a category is difficult under uncertain weakly supervised conditions.

Purpose of the Study:

• To propose a novel actionness-guided framework for suppressing false positives in WS-TAL.
• To improve the accuracy of temporal action localization by reducing background noise.
• To avoid introducing a separate background category, simplifying the learning process.

Main Methods:

• A self-training actionness branch learns class-agnostic actionness, minimizing label interference.
• A false positive suppression module identifies and removes false positive snippets.
• A foreground enhancement module utilizes attention and actionness to focus on relevant foreground actions.

Main Results:

• The proposed framework effectively suppresses false positives without explicit background modeling.
• Extensive experiments on THUMOS14, ActivityNet1.2, and ActivityNet1.3 demonstrate significant improvements.
• The method achieves state-of-the-art performance in weakly supervised temporal action localization.

Conclusions:

• The actionness-guided framework offers an effective solution for false positive suppression in WS-TAL.
• The approach enhances foreground action learning and localization accuracy.
• This work advances the field by providing a more robust and accurate WS-TAL method.