3.1. Datasets
3.1.1. Existing Datasets for Emotion Recognition
Emotion recognition research has greatly benefited froom datasets that combine EEG, ECG and other physiological signals to capture emotional response. There are many different datasets available, each with unique characteristcs, data structure and collection methods.
3.1.2. Overview of Common EEG/ECG Datasets for Emotion Recognition
Datasets like SEED , DEAP , and MAHNOB-HCI primarily use EEG combined with other signals such as eye tracking and physiological measures, and focus on emotions induced through videos or music clips in structured lab settings. These datasets typically label emotions based on dimensions like arousal, valence, and sometimes dominance or liking, offering extensive annotations for controlled stimuli. Other datasets like DERCFF and MPED expand further by integrating cardiac features and respiration signals to capture physiological responses across diverse environments, while DENS and EEWD emphasize naturalistic and real-time user interactions, reflecting more unstructured emotional responses. THU-EP presents a more experimental approach, focusing solely on EEG signals with unique, less conventional emotional labels.
DREAMER and AMIGOS datasets, on the other hand, include both EEG and ECG signals (with DREAMER using film clips and AMIGOS also incorporating group settings), allowing for more exploration of brain-heart interactions.
Not all these datasets include both EEG and ECG recorded together, however, they can contribute valuable insights and foundational models that can be adapted or or transferred to datasets like DREAMER or AMIGOS, which does include simultaneous EEG-ECG data.
The Table 2.1 below provides an overview of some commonly used emotion recognition datasets, comparing their primary collected signals, emotion labels, dataset structure, and how the data was collected.
2.1: Overview of Commonly Used EEG/ECG Datasets for Emotion Recognition
SEED
EEG, Eye Tracking, Facial Expression
Positive, Negative, Neutral
EEG signals from 15 subjects over 62 video clips; labeled by emotion
EEG recorded during viewing of film clips; eye tracking, expression capture
DEAP
EEG, Peripheral Physiological Signals
Arousal, Valence, Dominance, Liking
32 channels, EEG, physiological data from 32 subjects; 40 music video clips
EEG and peripheral data recorded while watching music videos
MAHNOB-HCI
EEG, Eye Gaze, Physiological Signals
Arousal, Valence
EEG, gaze, physiological data from 30 participants; extensive labeling
EEG, eye tracking, and physiological measures recorded during emotion-inducing stimuli
DREAMER
EEG, ECG
Arousal, Valence, Dominance
14 EEG channels, ECG data, 23 subjects; labeled by arousal, valence, dominance
Film clips used to elicit emotions; ECG and EEG captured simultaneously
AMIGOS
EEG, ECG, GSR
Arousal, Valence, Dominance, Liking
EEG, ECG, GSR from 40 participants; individual and group sessions
Recorded in lab settings; includes individual and social interactions
DERCFF [7
Cardiac Frequency, Facial Features
Arousal, Valence (inferred through physiology)
Cardiac frequency and facial features recorded in multiple settings
Recorded in various settings using wearable devices
DENS
EEG, Physiological Signals
Naturalistic emotional responses
EEG, physiological signals; unstructured naturalistic sessions
Naturalistic data collection with EEG and physiological monitoring
MPED
EEG, ECG, Respiration
Arousal, Valence (based on video stimuli)
EEG, ECG, respiration data from 30 subjects; based on emotional videos
Recorded during emotion-inducing videos in a lab environment
THU-EP
EEG
Experimental emotion labels
EEG data from 20 subjects; structured experimental protocol
Experimental setup using specific stimuli and EEG equipment
EEWD
EEG, Eye Tracking
Emotion-based user responses
EEG and eye tracking; recorded in dynamic user interaction environments
Recorded during real-time user interactions with wearable EEG and eye tracking
3.1.3. Rationale for Using DREAMER Dataset
Datasets such as DERCFF, THU-EP, EEWD are not publicly available. As shown in Table 2.1, other datasets such as MPED, DEAP, SEED and AMIGOS include a wide range of signals with detailed emotional labeling, giving comprehensive insights for various emotion recognition tasks. However, accessing many of these datasets, like DEAP and SEAD, often requires insitutional affliation or supervisor endorsement. Also, they can be technically demanding, especially in terms of preprocessing and handling complexity of multimodal signals.
Given my non-STEM background, choosing a simpler, more accessible dataset is my priority. This allows me to focus on emotion recognition task without getting overwhelmemed by technical challenges. DREAMER stands out as an ideal choice. It has a structured format, clear emotional labelling and straightforward access requirements. I only needed to submit a written request to the authors for access, making it practical for independent study.
The DREAMER dataset includes EEG and ECG recordings from 23 participants, captured as they respond to emotionally charged film clips. Participants also rated their emotions based on arousal, valence, and dominance dimensions. By focusing on DREAMER, this project can focus on exploration of brain-heart interactions in a manageable way, allowing for effective analysis of multimodal data without excessive complexity.
Last updated