Task-aware multiple instance learning for stress detection from facial video data.

Area of Science:

• Psychology
• Computer Science
• Biomedical Engineering

Background:

• Stress is a widespread issue impacting mental and physical health.
• Early stress detection is crucial for timely intervention and well-being.
• Current physiological stress detection methods are often intrusive and not scalable.

Purpose of the Study:

• To propose a non-intrusive, video-based stress detection method.
• To leverage Multiple Instance Learning for improved stress detection accuracy.
• To develop a method suitable for continuous and large-scale stress monitoring.

Main Methods:

• Utilized top-k Multiple Instance Learning with a temporal feature network and multi-head attention.
• Incorporated a conditioning mechanism for active and passive tasks.
• Included both top-k and bottom-k instances to utilize limited datasets and weak labels.

Main Results:

• Achieved high accuracy and F1 scores on custom and public datasets (StressID).
• Outperformed baseline methods in a leave-five-subjects-out evaluation.
• Demonstrated interpretable temporal localization of stress-indicative behaviors.

Conclusions:

• The proposed video-based method offers an effective and non-intrusive approach to stress detection.
• The top-k Multiple Instance Learning framework enhances accuracy by considering diverse behavioral patterns.
• This method holds promise for scalable and continuous stress monitoring applications.