---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: conditioning_image
    dtype:
      array2_d:
        shape:
        - 128
        - 440
        dtype: float32
  - name: caption
    dtype: string
  - name: label_folder
    dtype: string
  - name: label
    dtype: int32
  - name: subject
    dtype: int32
  splits:
  - name: train
    num_bytes: 2518963666.125
    num_examples: 7959
  - name: test
    num_bytes: 626837215.625
    num_examples: 1987
  - name: validation
    num_bytes: 630326693.75
    num_examples: 1994
  download_size: 3410667247
  dataset_size: 3776127575.5
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
language:
- en
tags:
- EEG
- Image
pretty_name: EEG-Based Visual Classification Dataset
---


# EEG-Based Visual Classification Dataset

This dataset is used in the paper "Guess What I Think: Streamlined EEG-to-Image Generation with Latent Diffusion Models" available on [arxiv](https://arxiv.org/abs/2410.02780) and the code is available at this [repository](https://github.com/LuigiSigillo/GWIT).

The dataset is constructed to be used in controlnet scenarios.

This dataset includes EEG data from 6 subjects. The recording protocol included 40 object classes with 50 images each, taken from the ImageNet dataset, giving a total of 2,000 images. Visual stimuli were presented to the users in a block-based setting, with images of each class shown consecutively in a single sequence. Each image was shown for 0.5 seconds. A 10-second black screen (during which we kept recording EEG data) was presented between class blocks. The collected dataset contains in total 11,964 segments (time intervals recording the response to each image); 36 have been excluded from the expected 6×2,000 = 12,000 segments due to low recording quality or subjects not looking at the screen, checked by using the eye movement data. Each EEG segment contains 128 channels, recorded for 0.5 seconds at 1 kHz sampling rate, represented as a 128×L matrix, with L about 500 being the number of samples contained in each segment on each channel. The exact duration of each signal may vary, so we discarded the first 20 samples (20 ms) to reduce interference from the previous image and then cut the signal to a common length of 440 samples (to account for signals with L < 500). The dataset includes data already filtered in three frequency ranges: 14-70Hz, 5-95Hz and 55-95Hz.

Download dataset Link to dataset files: https://tinyurl.com/eeg-visual-classification

```python
from datasets import load_dataset
split = "train" #could be one of "train", "test", "validation"
data = load_dataset('luigi-s/EEG_Image_CVPR_ALL_subj', split=split).with_format(type='torch')
eeg = data[0]['conditioning_image']
image = data[0]['image']
caption = data[0]['caption'])
subject = data[0]['subject'])
label =   data[0]['label'])
```