Uncertainty-Aware Dual-Evidential Learning for Weakly-Supervised Temporal Action Localization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Weakly-supervised temporal action localization (WTAL) aims to identify action instances and categories using only video-level labels.
• Existing WTAL methods struggle with action-background ambiguity due to background noise and overlooking non-salient action segments.

Purpose of the Study:

• To propose a novel framework, Uncertainty-aware Dual-Evidential Learning (UDEL), to address action-background ambiguity in WTAL.
• To leverage epistemic and aleatoric uncertainties within an evidential deep learning (EDL) framework for improved performance.

Main Methods:

• UDEL fuses video-level epistemic and aleatoric uncertainties to quantify background noise interference.
• Snippet-level aleatoric uncertainty is deduced for progressive mutual learning, focusing on action instances in an "easy-to-hard" manner.
• The framework encourages snippet-level epistemic uncertainty to complement foreground attention scores.

Main Results:

• UDEL achieves state-of-the-art performance on four public benchmarks.
• The method effectively reduces action-background ambiguity.
• Experimental results demonstrate the efficacy of the proposed uncertainty-guided approach.

Conclusions:

• UDEL provides a robust solution for weakly-supervised temporal action localization.
• The integration of dual uncertainties enhances the model's ability to handle noisy and complex video data.
• The proposed framework offers a significant advancement in the field of action localization.