Datasets:
EEG Categorical Object Repetition Dataset (EEG-CORD)
EEG-CORD is an EEG dataset designed to study how experimental protocol choices influence multiclass visual object decoding. The dataset was collected to support controlled analyses of block structure, within-class repetition, and session effects in EEG-based object recognition.
EEG-CORD is intentionally a methodological dataset rather than a leaderboard benchmark. Its design emphasizes interpretability, control of confounds, and evaluation protocol analysis.
Dataset overview
| Property | Value |
|---|---|
| Modality | EEG |
| Participants | 11 |
| Sessions per participant | 3 |
| Object categories | 10 |
| Stimulus type | AI-generated object images |
| Trial structure | Short class-specific blocks |
| Images per block | 4 standard images |
| Jitter trials | Optional fifth image |
| Primary focus | Block design, repetition, session progression |
What this dataset is for
EEG-CORD was created to make common EEG decoding design choices measurable under controlled conditions. In particular, it supports research on:
- whether short block-based presentation introduces evaluation artifacts (temporal leakage) when trials in blocks are split incorrectly,
- whether repetition improves decoding when temporal leakage is prevented,
- how decoding changes over the course of a session.
The dataset is not intended as a large-scale natural-image EEG benchmark. Its purpose is to isolate methodological effects in a compact and reproducible setting.
Experimental design
Participants viewed images from ten object categories:
- airplane
- apple
- banana
- boat
- car
- cat
- clock
- dog
- rocket
- train
Stimuli were presented in short blocks of four images from the same class. Blocks were randomized, and blocks from the same class were never presented consecutively.
Each block contained:
- a first image,
- two middle images,
- a final image,
with an optional fifth jitter image presented with 50% probability.
Images were shown for 1 second each. Inter-stimulus and inter-block intervals (between 0.75s and 1.25s, and between 4.5s and 5.5s) were jittered. Participants received a 20s break after every tenth block. Each image was shown only once per participant, which prevents repeated-stimulus leakage across splits.
Each session comprised 9 × 10 (i.e., 90) blocks.
Stimulus coding
Each stimulus onset is marked by a two-digit event code that encodes both the object category and the position of the image within its block.
The first digit corresponds to the class ID (
0–9).The second digit indicates the image's position within the block:
1- first image2- second or third image (the two middle images)3- fourth (last) image9- optional jitter image
For example, event code 42 denotes one of the two middle images in a car block (class_id = 4), while 73 denotes the fourth (last) image in a dog block (class_id = 7). Single-digit values are written with a leading zero (e.g., 1 is stored as 01).
At the beginning of each recording session, two 30-second resting-state recordings are acquired before the stimulus presentation begins: first eyes closed, followed by eyes open. Event code 150 marks the auditory cue instructing participants to open their eyes, separating the two resting-state recordings.
Data collection
Data were collected across three sessions per participant, with sessions separated by at least 24 hours.
Participants
Twelve participants were recruited initially. One withdrew after the first session, leaving data from 11 participants. All participants had normal or corrected-to-normal vision.
EEG recording
EEG was recorded with a 64-channel Quik-Cap Hydro Net connected to a SynAmps RT amplifier and controlled via Curry 8. Signals were sampled at 1 kHz with 24-bit resolution. Stimulus presentation was controlled using STIM2, with event codes synchronized to the EEG recording system.
Stimuli
All stimuli were AI-generated to maintain tight control over visual content and to minimize irrelevant background cues. Each image depicts a single isolated object. The stimuli were generated with black-forest-labs/FLUX.1-dev and manually screened to remove implausible or low-quality images.
File structure
The dataset is organized as follows:
raws/
├── subject_{subj:02d}_session_{session}_raw.fif
└── sequences.h5
The .fif recordings and the sequence file are stored in /raws.
Image folders
Images are stored by category and by image type.
Standard images
images/airplane_0_images/
This folder contains the standard airplane images, with filenames such as:
airplane_0.jpg
airplane_1.jpg
airplane_2.jpg
...
Here, 0 is the class_id for airplane, and the number in the filename is the image_id.
Image ID numbering across categories
The image_id numbering continues across classes. Each class contains 108 images, so the next class begins where the previous one ended.
For example, in:
images/apple_1_images/
the first apple image is named:
apple_108.jpg
because airplane images occupy 0–107, and apple images continue from there.
Jitter images
Jitter images are stored separately, for example:
images/airplane_0_jitter_images/
This folder contains files such as:
airplane_jitter_0.jpg
Metadata
Trial metadata are stored in the raws/sequences.h5 file. It contains the following structure:
['sequences']['subject_{subj:02d}']['session_{session}']['csv']
Metadata columns
sequence_index: block number within the session, starting from 1class_name: stimulus class nameclass_id: numeric class identifiertype:standardorjitterstim_code: event marker sent to the EEG systemimage_id: unique image identifierimage_duration_ms: duration of the image presentationsequence_duration_ms: duration of the full block, defined for the first image in each block only
Intended use
EEG-CORD is suitable for work on:
- block-based EEG decoding,
- repetition and leakage confounds,
- within-session temporal effects,
- protocol-sensitive benchmarking,
- preprocessing sensitivity analysis,
- representational analysis of object-related EEG responses.
Limitations
EEG-CORD was designed for control and interpretability, not for scale.
Key limitations include:
- a relatively small number of participants,
- a limited number of categories,
- AI-generated stimuli rather than natural photographs,
- no auxiliary attention task during stimulus presentation.
These properties make the dataset especially useful for protocol studies, but less suitable as a broad naturalistic benchmark.
Ethical considerations
The study was approved by the University of Otago Human Ethics Committee (Health) [H25/0492]. Participants provided informed consent, could withdraw at any time during the collection, and all released data are anonymized.
Citation
Please cite the associated paper when using this dataset (paper in preparation).
@article{EEGCORD,
title = {EEG Categorical Object Repetition Dataset (EEG-CORD)},
author = {Heinrich, Lukas and Liesaputra, Veronica and Huang, Zhiyi},
year = {2026}
}
- Downloads last month
- 3