Vectorized Evidential Learning for Weakly-Supervised Temporal Action Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Analysis

Background:

• Weakly-supervised temporal action localization (WS-TAL) is crucial for video analysis, but faces challenges from weak supervision and open-world scenarios.
• Existing deep learning methods show progress but struggle with robustness and reliability due to inherent uncertainties.

Purpose of the Study:

• To introduce a novel paradigm, Vectorized Evidential Learning (VEL), for improved WS-TAL.
• To address challenges of uncertainty, noisy data, and unknown categories in WS-TAL.

Main Methods:

• VEL utilizes learnable meta-action units (MAUs) as fundamental building blocks for action categories.
• It dynamically learns action components and their relations using MAUs and category representations, incorporating uncertainty estimation.
• Local evidence is accumulated and optimized using Subject Logic theory.

Main Results:

• VEL demonstrated consistent improvements in robust and reliable action localization across regular, noisy, and open-set settings.
• Experiments on three benchmarks confirmed VEL's superior performance compared to state-of-the-art methods.

Conclusions:

• Vectorized Evidential Learning offers a promising new direction for WS-TAL.
• The proposed method effectively handles uncertainties and improves localization accuracy in challenging video analysis tasks.