Compact Representation and Reliable Classification Learning for Point-Level Weakly-Supervised Action Localization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Point-level weakly-supervised temporal action localization (P-WSTAL) uses single point labels for training.
• Existing localization-by-classification models suffer from large intra-action variation and noisy classification learning due to sparse annotations.

Purpose of the Study:

• To propose a novel framework, CRRC-Net, to address limitations in P-WSTAL.
• To improve the accuracy and reliability of temporal action localization and classification.

Main Methods:

• CRRC-Net employs a co-supervised feature learning module to leverage complementary information across modalities.
• A probabilistic pseudo label mining module estimates pseudo-sample likelihood using feature distances from action prototypes.
• This module rectifies pseudo-labels for more reliable classification learning.

Main Results:

• The proposed CRRC-Net framework effectively addresses intra-action variation and noisy classification learning.
• Experiments on benchmark datasets demonstrate favorable performance compared to state-of-the-art methods.
• The co-supervised and probabilistic modules contribute to enhanced feature representations and reliable learning.

Conclusions:

• CRRC-Net offers a robust solution for P-WSTAL by tackling key challenges in existing methods.
• The framework achieves state-of-the-art results, highlighting the effectiveness of its novel modules.
• This work advances the field of weakly-supervised action localization with improved accuracy and reliability.