PAIR: Perception And Imagery Recall

EEG dataset for Replaying the Movie in Your Mind: Decoding Dynamic Visual Perception and Recall from EEG, accepted at ICONIP 2026.

Code and archived-score reproduction: freesky3/PAIR. Original code in that repository is MIT-licensed; historical and third-party files retain their own terms. Data version: 1.0.0. License: CC BY-NC 4.0, allowing attribution-based noncommercial use and adaptation. Commercial use requires separate permission from the rights holders.

Contents and scale

Twenty participants (10 male and 10 female, as reported in the manuscript) completed three recording sessions each. Each session contains five blocks of 50 trials: 250 unique two-second videos, hierarchically labeled into four broad and twenty fine categories. Repeated presentations produce 15,000 paired trials (30,000 perception/recall epochs), not 15,000 unique videos.

The trial sequence was: 2 s viewing, 2 s recall-initiation prompt, 3 s recall, 3 s termination/break. Each block started with a 5 s readiness prompt. EEG was acquired at 1,000 Hz using 62 channels and downsampled to 200 Hz after filtering and ICA artifact attenuation. Eye tracking was recorded but not analyzed in the paper and is not included here. Unprocessed raw acquisition files and eye-tracking records are not included. Experimental stimulus video clips and task cue media are included under the ideos/ directory.

Directory	Single recording shape	Axes
watch_cleaned	(5, 50, 62, 400)	block, trial, channel, time
recall_cleaned	(5, 50, 62, 600)	block, trial, channel, time
watch_PSD_DE	(2, 5, 50, 62, 5)	PSD/DE, block, trial, channel, band
recall_PSD_DE	(2, 5, 50, 62, 5)	PSD/DE, block, trial, channel, band

The feature axis is PSD first, DE second. Frequency bands described in the manuscript are 1–4, 4–8, 8–14, 14–31 and 31–99 Hz. The cleaned arrays are preprocessed time-domain EEG. Values and dtypes are preserved byte-for-byte from the authors' supplied arrays; no additional amplitude scaling has been applied.

Video stimuli and task cues

The stimulus videos and experimental media corresponding to the EEG recordings are organized under ideos/:

• ideos/shuffled_renamed_videos/: Contains the 250 unique 2-second stimulus clips presented across the 5 experimental blocks (group1/ to group5/, with 50 clips per group, named 1.mp4 through 50.mp4). These correspond to the block and trial ordering in the EEG data and align with metadata/stimulus_labels.csv.
• ideos/groupedVideos/: Contains 5 stitched video streams (output_group_1.mp4 through output_group_5.mp4), each concatenating the entire sequence for one block as presented during the experiment.
• ideos/attachment/: Contains task instruction, transition, and cue media used in the experimental pipeline (start.mp4, lank.mp4, prompt_play.mp4, prompt_recall.mp4, hank_you_for_watching.mp4, and ding.mp3).

Metadata and identity

• recordings.csv preserves the original recording-processing order using source_index and anonymous IDs. recording-01 through recording-03 denote within-participant sorted order, not the literal session labels in the original filenames.
• metadata/stimulus_labels.csv contains all 250 clip labels in flattened block/trial order. It is safe to read without NumPy pickle loading.
• metadata/channels.tsv lists the 62 electrode names in array order. These names do not establish source localization.
• metadata/adj_matrix.npy contains the supplied electrode graph; other numeric metadata reproduce original semantic/flow labels.
• manifest.json lists hashes, sizes, shapes and dtypes for the numeric files and metadata.

Original personal filenames and the private identity mapping are not distributed. IDs align with the anonymous result tables in the code repository. Written informed consent is reported in the manuscript; no additional approval identifier is asserted by this data card.

Label definitions

Fine semantic labels in GT_label.npy are 1–20; coarse labels are (label - 1) // 5. Fast/slow uses a float32 optical-flow comparison with threshold 0.6427. Color classes are Neutral Light, Earth & Dark, Cool Tones, Green Nature and Warm Vibrant. Number labels are 0 for up to one object, 1 for two through four, and 2 for more than four. Face/human labels indicate presence.

Full-stimulus majority-class proportions are 6.0% (20-c), 25.2% (4-c), 50.0% (Fast/Slow), 22.8% (Color), 43.6% (Number), 60.4% (Face), and 54.4% (Human). These describe label distribution, not the accuracy of a historical validation-set baseline.

Download and validation

Large EEG arrays are stored as lossless parts of at most 4 MiB to support reliable resumption over interrupted connections. transport.json specifies the ordered parts and their SHA-256 hashes. The PAIR download command automatically restores the original .npy files and verifies that the reconstructed bytes match manifest.json. PSD/DE files and previously uploaded waveforms remain whole. Allow about 15 GB of local space while keeping both parts and restored arrays.

After cloning the code repository and running uv sync --locked:

uv run pair-eeg download --repo skywalker-p/PAIR --revision v1.0.0 --data-dir data
uv run pair-eeg validate-data --data-dir data

A direct Hub download retrieves transport parts. Assemble them using the PAIR tool:

hf download skywalker-p/PAIR --repo-type dataset --revision v1.0.0 --local-dir data
uv run pair-eeg assemble-data --data-dir data
uv run pair-eeg validate-data --data-dir data

import numpy as np
watch = np.load("data/watch_cleaned/sub-001_recording-01.npy", allow_pickle=False)
epochs = watch.reshape(250, 62, 400)

Evaluation scope and limitations

The archived benchmark fits models separately for each participant/session. Content decoding randomly partitions 250 epochs into 200 training and 50 validation samples. State decoding uses the first 200 clips per condition for training and the last 50 for validation, keeping the two conditions of a clip together. Neural models report their best validation accuracy selected by early stopping; there is no independent test set. Exact historical split indices were not saved.

Content decoding uses different epoch durations (2 s perception and 3 s recall). Raw state decoding crops recall to 0.5–2.5 s. The archived raw alpha ablation removes 8–13 Hz, while spectral ablation omits the 8–14 Hz feature. No cross-subject evaluation or eye-tracking-based control was performed. Residual ocular signals, visual input and task differences may contribute to state discrimination. Topographic maps are descriptive sensor-level scores, not cortical source estimates.

The original reference-electrode details, physical amplitude units of the stored arrays, and complete raw-to-feature preprocessing script were not available in the supplied release materials. They are not inferred here. The manuscript reports 0.1–100 Hz filtering and ICA, but the full preprocessing provenance should be supplemented when the original records become available. Array integrity checks alone do not independently establish event/label alignment.

Citation and maintenance

Please cite Replaying the Movie in Your Mind: Decoding Dynamic Visual Perception and Recall from EEG (ICONIP 2026) and this dataset revision. The final proceedings DOI will be added when available. Report questions or issues through the PAIR code repository.