KFC: An Efficient Framework for Semi-Supervised Temporal Action Localization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised learning (SSL) is crucial for temporal action localization (TAL) to reduce annotation costs.
• Consistency regularization (CR) is effective in image classification but not directly applicable to TAL due to video length.
• Existing CR methods rely on input perturbations unsuitable for TAL's temporal nature.

Purpose of the Study:

• To develop a novel perturbation strategy for semi-supervised temporal action localization.
• To adapt consistency regularization for TAL by creating feature-level perturbations.
• To improve the robustness and accuracy of temporal action localization models.

Main Methods:

• Introduced K-farthest crossover, a new method for constructing perturbations based on video features.
• Applied CR by adding temporal perturbations to video features.
• Recombined features with perturbations to maintain semantics while introducing discrepancy, inspired by chromosomal crossover.

Main Results:

• The proposed K-farthest crossover method effectively generates suitable perturbations for TAL.
• Consistency regularization with these feature-based perturbations enhances model robustness.
• The approach encourages feature similarity within action instances and dissimilarity between different instances/backgrounds.

Conclusions:

• K-farthest crossover is a viable and effective perturbation technique for semi-supervised TAL.
• This method successfully adapts CR principles to the temporal domain of video analysis.
• The approach offers a promising direction for improving temporal action localization performance.