Dynamic Cascades with Bidirectional Bootstrapping for Action Unit Detection in Spontaneous Facial Behavior

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Automatic facial action unit (AU) detection is crucial for facial expression analysis.
• Existing methods often struggle with imbalanced datasets and high false positive rates due to suboptimal training image selection.
• Previous approaches include using peak frames (leading to imbalance) or all frames from onset to offset (increasing false positives).

Purpose of the Study:

• To propose a novel method for selecting training samples for automatic facial action unit detection.
• To address the limitations of existing training data selection strategies.
• To improve the accuracy and robustness of facial expression analysis systems.

Main Methods:

• Introduced dynamic cascades with bidirectional bootstrapping (DCBB) for optimal training sample selection.
• Utilized an iterative approach to select both positive and negative samples.
• Employed Cascade Adaboost as the base classifier, leveraging its feature selection, efficiency, and robustness.

Main Results:

• DCBB demonstrated improved AU detection performance compared to alternative methods on the RU-FACS (M3) database.
• The method effectively optimized the selection of positive and negative training samples.
• Improvements were observed for most tested action units in real-world, non-posed behavioral data.

Conclusions:

• DCBB offers a superior strategy for training sample selection in facial action unit detection.
• The proposed method enhances the accuracy of facial expression analysis, especially for challenging datasets.
• This approach contributes to more reliable automatic analysis of human emotions and behaviors.