A performance comparison of eight commercially available automatic classifiers for facial affect recognition.

Area of Science:

• Computer Science
• Psychology
• Affective Computing

Background:

• Commercial automatic facial affect recognition classifiers are advancing rapidly.
• Limited understanding exists regarding the comparative performance of these classifiers, especially for spontaneous facial expressions.

Purpose of the Study:

• To evaluate the performance of eight commercial automatic classifiers for facial emotion recognition.
• To compare the accuracy of these classifiers against human observers.
• To investigate performance differences between posed and spontaneous facial expressions.

Main Methods:

• Tested eight out-of-the-box automatic classifiers on 937 videos from BU-4DFE (posed) and UT-Dallas (spontaneous) databases.
• Videos depicted six basic emotions: happiness, sadness, anger, fear, surprise, and disgust.
• Compared classifier performance to human observer accuracy.

Main Results:

• Human observers significantly outperformed automatic classifiers in emotion recognition.
• Automatic classifier accuracy varied widely (48%-62%).
• Top classifiers achieved human-level accuracy for posed expressions but showed lower accuracy for spontaneous emotions.

Conclusions:

• Existing automatic classifiers show limitations in recognizing spontaneous emotions.
• There is a need for more diverse, spontaneous facial expression databases for training and testing.
• Current systems may be less reliable for real-world emotion analysis outside controlled settings.