Dataset Card for EEG-ImageNet

Dataset Description

Exploring brain activity in relation to visual perception provides insights into the biological representation of the world. While Functional Magnetic Resonance Imaging (fMRI) and magnetoencephalography (MEG) have enabled effective image classification and reconstruction, their high cost and bulk limit practical use. In contrast, Electroencephalography (EEG) offers advantages like affordability and high temporal resolution, but is underutilized due to a lack of comprehensive datasets. To fill this gap, we introduce EEG-ImageNet, a novel EEG dataset featuring recordings from 16 participants, with a total of 22 sessions (i.e., 6 participants took part in two sessions each), exposed to 4000 images sourced from the ImageNet dataset. EEG-ImageNet includes image stimuli labeled with varying levels of granularity, comprising 40 images with coarse labels and 40 with fine labels. We establish benchmarks for both object classification and image reconstruction based on this dataset.

Dataset Structure

Figure 1: Schematic Diagram of the Data File Storage Structure.

Due to file size limitations on the cloud storage platform, the dataset is split into two parts: EEG-ImageNet_1 and EEG-ImageNet_2. Each part contains data from 8 participants. Users can choose to use only one part based on their specific needs or device limitations.

The EEG-ImageNet dataset contains a total of 87,850 EEG-image pairs from 16 participants. Each EEG data sample has a size of (n_channels, $f_s \cdot T$), where n_channels is the number of EEG electrodes, which is 62 in our dataset; $f_s$ is the sampling frequency of the device, which is 1000 Hz in our dataset; and T is the time window size, which in our dataset is the duration of the image stimulus presentation, i.e., 0.5 seconds. Due to ImageNet's copyright restrictions, our dataset only provides the file index of each image in ImageNet and the wnid of its category corresponding to each EEG segment.

Figure 2: The overall procedure of our dataset construction and benchmark design. The experimental paradigm involves four stages: S1: Category Presentation (displaying the category label), S2: Fixation (500 ms), S3: Image Presentation (each image displayed for 500 ms), and S4: an optional random test to verify participant engagement. Each image presentation sequence includes 50 images from the given category, during which EEG signals are recorded. Data flow is indicated by blue arrows, while collected data is highlighted in gray. The stimuli images are sourced from ImageNet, with EEG signals aligned to image indices, granularity levels, and labels. The benchmarks (image reconstruction and object classification) are designed to evaluate coarse and fine granularities classification tasks.

Table 1: The average results of all participants in the object classification task. * indicates the use of time-domain features, otherwise it indicates the use of frequency-domain features. † indicates that the difference compared to the best-performing model is significant with p-value < 0.05.

Model		Acc (all)	Acc (coarse)	Acc (fine)
Classic model	Ridge	0.286±0.074 †	0.394±0.081 †	0.583±0.074 †
	KNN	0.304±0.086 †	0.401±0.097 †	0.696±0.068 †
	RandomForest	0.349±0.087 †	0.454±0.105 †	0.729±0.072 †
	SVM	0.392±0.086	0.506±0.099 †	0.778±0.054 †
Deep model	MLP	0.404±0.103	0.534±0.115	0.816±0.054
	EEGNet*	0.260±0.098 †	0.303±0.108 †	0.365±0.095 †
	RGNN	0.405±0.095	0.470±0.092 †	0.706±0.073 †

Figure 3: The image reconstruction results of a single participant (S8).