THINGS: a multimodal dataset for investigating object representations in thehuman brain

It’s actually a collection of 3 datasets. All of which can be found here. The images used can be found here.

The 3 datasets are: 1. FMRI scans from 3 participants. 8740 images. 2. MEG scans. 4 participants. 22k images. 3. 12k participants. 4.7 million similarity judgements.

The 2 main datasets I’m interested in are the MEG scans and the similarity judgements. For the MEG scans, we can use these scripts as boilerplate.

For now lets stick to MEG as planned.

Notes on MEG data collection: - Each session consisted of 10 runs (~5 min each). - In each run, 185–186 object images were presented, as well as 20 test and 20 catch images - Stimuli were presented for 500ms - 225–226 trials per run, 2,254 trials per session - 27,048 trials per participant - Stimulus presentation was done with Psychtoolbox

Metadata for a single MEG run for session 1 run 9:

{
    "TaskName": "main",
    "Manufacturer": "CTF",
    "PowerLineFrequency": 60,
    "SamplingFrequency": 1200.0,
    "SoftwareFilters": "n/a",
    "RecordingDuration": 347.99916666666667,
    "RecordingType": "continuous",
    "DewarPosition": "n/a",
    "DigitizedLandmarks": false,
    "DigitizedHeadPoints": false,
    "MEGChannelCount": 272,
    "MEGREFChannelCount": 28,
    "EEGChannelCount": 0,
    "EOGChannelCount": 0,
    "ECGChannelCount":+ 0,
    "EMGChannelCount": 0,
    "MiscChannelCount": 10,
    "TriggerChannelCount": 0
}

I think "MEGChannelCount": 272 refers to the fact that we have 272 channels. So if we average along the time dimension, then for each instance, we get a tensor of shape 272.

Files like these contain the index of the image used (most probably it’s that)

Open questions: 1. what is TRIAL_TYPE? (values: ['exp', 'test', 'catch'...])