Active Visual Semantics

The Active Visual Semantics (AVS) Dataset is a multimodal neuroimaging dataset combining magnetoencephalography (MEG), eye-tracking, and structural MRI, recorded from 5 participants (sub-01-sub-05) as they actively explored 4,080 natural scenes (subsampled from the Natural Scenes Dataset, NSD) across 10 recording sessions each, yielding more than 200,000 fixation epochs in total. Unlike neuroimaging datasets that rely on passive viewing with enforced central fixation, AVS captures brain activity during active, self-directed scene exploration, including natural saccades and fixations. A semantic scene-captioning task on 25% of trials links gaze behaviour to scene understanding and memory. For each session we provide eye-movement-locked MEG epochs (fixation- and saccade-locked; epochs for other events such as scene onset can be easily recomputed), raw and preprocessed eye-tracking data, per-fixation object category labels, human ratings of whether fixation targets were mentioned in participants' scene captions, pupil dynamics, and defaced structural MRI scans. Individualised head stabilisation casts, together with the structural scans, enable precise source reconstruction at the single-participant level across sessions. Source-reconstruction derivatives (FreeSurfer cortical surfaces, BEM models, forward solutions) are provided for every subject and are directly usable as an MNE-Python SUBJECTS_DIR. The dataset is accompanied by the open-source pyAVS Python package for loading, preprocessing, and analysis.