Knowledge Conditioned Variational Learning for One-Class Facial Expression Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial Expression Recognition (FER) systems struggle with diverse, real-world expressions due to data limitations.
• Detecting 'alien' expressions (unseen during training) is crucial for robust FER systems.

Purpose of the Study:

• To develop a novel FER model, OneExpressNet, capable of quantifying the probability of a test expression belonging to the training data distribution.
• To enhance the robustness of FER systems by effectively identifying and handling expressions absent during training.

Main Methods:

• Proposed OneExpressNet model based on a variational auto-encoder.
• Incorporated knowledge transfer from related domains as a constraint for target distribution.
• Utilized skip connections between encoder and decoder for feature aggregation.
• Employed empirical compact variation loss to concentrate training expressions on a compact manifold.

Main Results:

• Achieved state-of-the-art results in one-class facial expression recognition.
• Demonstrated effectiveness on both small-scale lab-controlled datasets (CFEE, KDEF) and large-scale in-the-wild datasets (RAF-DB, ExpW).

Conclusions:

• OneExpressNet effectively quantifies the probability of test expressions belonging to the training distribution.
• The model enhances FER system robustness by addressing the challenge of alien expressions.