Hugging Face's logo

Datasets:
Fudan-fMRI
/
CineBrain

Tasks:
Text-to-Video
Image-to-Video
Modalities:
Text
Formats:
webdataset
Languages:
English
Size:
100K - 1M
ArXiv:
Tags:
fMRI
EEG
neuroscience
brain-imaging
science
huggingscience
Libraries:
Datasets
WebDataset
License:
Dataset card Data Studio Files
xet
Community
4

Acknowledgment and Questions on the CineBrain Dataset

#4
by HopkinsKwong - opened
Discussion
HopkinsKwong

Dear Authors,

I am writing to express my sincere gratitude for your outstanding work on the CineBrain dataset, which has provided invaluable support for my research on EEG-based multimodal neural decoding.

I have encountered several questions while using the dataset (https://huggingface.co/datasets/Fudan-fMRI/CineBrain/tree/main) and would greatly appreciate your clarification:

  1. The preprocessed EEG data has a shape of (69, 800). Since the original EEG was collected with 64 channels, are the first 64 rows EEG signals and the last 5 rows ECG signals?

  2. Additionally, does this shape imply downsampling from 1000 Hz to 200 Hz (800 data points for 4 seconds)?

  3. The EEG files for sub0002 and sub0003 appear identical. Could this be a data upload error?

  4. The video data lacks corresponding audio files. Would it be possible to provide the audio data for alignment with EEG signals?

  5. Regarding data splits: Is there a formal table mapping clips to specific episodes? We hypothesize 270 clips per episode (5400 clips/20 episodes), but confirmation is needed.

  6. Each subject has 27000 EEG npy files, while there are only 5400 video/audio clips. What is the exact correspondence between these EEG files and the 5400 clips?

Thank you again for your contributions and for your time to address these questions. I look forward to your reply.

Best regards.
Jianxiong
 Fudan-fMRI-yanwei org
•
edited 26 days ago

Thanks for your interest in our work. We address your questions point by point below:

  1. Yes — the first 64 channels correspond to EEG signals, and the last channel is ECG.
  2. During preprocessing, EEG and fMRI are temporally aligned. fMRI is sampled every 0.8 s, and EEG is segmented accordingly into 0.8 s windows, resulting in 800 data points per segment. The original EEG sampling rate remains 1000 Hz.
  3. After checking, we confirm that there were indeed some issues in the currently uploaded data. These have now been corrected.
  4. The audio files were previously uploaded to ModelScope and have not yet been synchronized to Hugging Face. They are available at: https://www.modelscope.cn/datasets/Jianxionggao/CineBrainVide.
    In that repository, both videos and audio are segmented into 4 s clips. The videos are at 16.25 fps, which is slightly higher than those on Hugging Face. The choice of version depends on your specific use case.
  5. For each episode, we use the first 18 minutes and segment them into 4 s clips, yielding 18 × 60 / 4 = 270 clips per episode.
    For videos, we select the first 10 episodes from Seasons 7, 9, and 11. Video IDs follow the order 0701, 0702, …, 0901, 0902, …, 1109, 1110 (season + episode index).
  6. Different subjects watched different sets of videos:
  • Subjects 1, 2, and 6 watched Seasons 7 and 9, corresponding to video IDs 1–20 (5400 clips).
  • Subjects 3, 4, and 5 watched Seasons 7 and 11, corresponding to video IDs 1–10 and 21–30 (5400 clips).
    As long as the video IDs are correctly matched, EEG and fMRI data correspond one-to-one.

If you have any further questions, please feel free to discuss them with us.

Best regards,
Jianxiong

Sign up or log in to comment