中文版

Experiment Setup

As shown in the figure below, the experiment was designed to investigate neural patterns of odor-induced emotional states. Participants completed three independent experimental sessions, with at least 24 hours between sessions to maintain data independence.
SEED-OLF experiment protocol
Each session consisted of six folds, and within each fold, four different odorants (two pleasant and two unpleasant) were presented once in a randomized sequence, resulting in 24 trials per session. Each trial followed a standardized procedure composed of four stages:
  1. Preparation (~1 min): The subject received visual instructions on an interactive display to prepare for the upcoming trial.
  2. Stimulation (15 s): Odor was delivered via an olfactory mask. Subjects were instructed to inhale normally.
  3. Self-assessment (~10 s): Subjects reported their perceived emotional states (pleasant or unpleasant) using the interface.
  4. Clean Breathing (30 s): Clean air was delivered to purge residual odor and allow the subject to return to baseline.

Subjects

A total of 32 healthy right-handed subjects (16 males and 16 females) aged from 18 to 28 years (mean age: 24.19) were recruited. All subjects had normal or corrected-to-normal vision and a normal sense of smell. They were screened to exclude respiratory diseases, olfactory sensitivity issues, or a history of neurological/psychiatric disorders.

Stimulus Material

Following a pilot study of 21 odorants, four representative odorants were selected for the formal experiments: pleasant odorants (high valence) #19 Green tea and #20 Orange, and unpleasant odorants (low valence) #5 Isovaleric acid and #9 3-methylindole. The odorants were delivered using a programmable air pump system through an olfactory mask to ensure precision and synchronization with EEG recording.
Candidate odorants

EEG Signals

EEG signals were recorded using the ESI NeuroScan system with a 62-channel cap (international 10-20 system) at a sampling rate of 1000 Hz. A band-pass filter of 0.1-70 Hz and a 50 Hz notch filter were applied to remove noise and power-line interference. The signals were downsampled to 200 Hz.

Classification Tasks

The SEED-OLF dataset is designed for three major tasks:
  1. SNSC Task: Sniffing vs. Non-Sniffing State Classification.
  2. OSEC Task: Olfactory-Induced Subjective Emotion Classification (Positive vs. Negative).
  3. OOEC Task: Olfactory-Induced Objective Emotion Classification (based on odor valence).

Dataset Summary

To ensure privacy, all participant names have been replaced with IDs (01 to 32). The dataset is organized into two main folders: stimulation and clean_breathing.
  1. Folder "stimulation": Contains EEG data from the 15-second stimulation phase.
  2. Folder "clean_breathing": Contains EEG data from clean breathing (rest) phase.
  3. File Naming: Data files follow the format subID_Session_Trial.pkl (e.g., 01_2_1.pkl indicates data from the first participant during the first trial of the second session).
  4. Data Format (.pkl): Each file contains a dictionary with X_raw, y_odor, y_emotion, and y_sniff.
  • X_raw: Raw EEG signals with a shape of (62, 3000) for stimulation or clean breathing.
  • y_odor: Odor label (1: Green tea, 2: Orange, 3: Isovaleric acid, 4: 3-methylindole).
  • y_emotion: Subjective emotion label (0: Positive, 1: Negative).
  • y_sniff: Sniffing status label (0: Non-sniffing, 1: Sniffing).

Download

Download SEED-OLF

Reference

If you feel that the dataset is helpful for your study, please add the following reference to your publications.

J. -M. Zhang, W. -B. Jiang, W. -L. Zheng and B. -L. Lu, "SEED-OLF: a Novel EEG Dataset with Olfactory Stimulation for Emotion Recognition," in IEEE Transactions on Affective Computing, doi: 10.1109/TAFFC.2026.3662364. [link] [BibTex]