K-EmoCon: Continuous Emotion Annotations from Three Perspectives During Debates

K-EmoCon is the first publicly available emotion dataset accommodating multiperspective assessment during social interactions. It contains multimodal sensor data from 16 paired debate sessions, each approximately 10 minutes long, with continuous emotion annotations made every 5 seconds. The dataset was created by Cheul Young Park at KAIST and released in 2020.

Use Cases

• Training models for continuous emotion recognition based on multimodal sensor streams including EEG and peripheral physiology.
• Studying discrepancies in emotional perception based on self, partner, and external observer annotations.
• Developing fusion techniques for audiovisual, physiological, and neural signals mentioned in the description.
• Analyzing arousal-valence dynamics during naturalistic conversations described in the dataset.

Strengths

• Includes emotion annotations from three distinct perspectives: self, debate partner, and external observers.
• Multimodal measurements encompass audiovisual recordings, EEG, and peripheral physiological signals.
• Annotations were made at a high temporal resolution of every 5 seconds during debates.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

Korea Advanced Institute of Science and Technology (KAIST)

Collection Method

Data was acquired with off-the-shelf devices from 16 sessions of paired debates on a social issue.