EgoBrain: Synergizing Minds and Eyes For Human Action Understanding
Paper • 2506.01353 • Published
[ICLR 2026]
EgoBrain: Synergizing Minds and Eyes For Human Action Understanding
EgoBrain-Mini — lightweight, fast, and fully synchronized
40 subjects · 1,562 sub-actions · 63 hours (61 + 2 calibration) · 4 synchronised modalities · 27 everyday activities
東京大学 The University of Tokyo X 微軟亞洲研究院 Microsoft Research Asia
Nie Lin · Yansen Wang · Dongqi Han · Weibang Jiang · Jingyuan Li · Ryosuke Furuta
Explore the project page →
The room above is live there — tap the desk objects, the EEG cap, the camera or the wrist bands.
This is EgoBrain-Mini, a lightweight companion release of our ICLR 2026 paper "EgoBrain: Synergizing Minds and Eyes For Human Action Understanding" — the egocentric subset (first-person video + EEG + IMU + audio), meant as a quick way to explore the data and its alignment before diving into the full EgoBrain release.
Nie (Elon) Lin, Yansen Wang, Dongqi Han, Weibang Jiang, Jingyuan Li, Ryosuke Furuta, Yoichi Sato*, Dongsheng Li* (*co-corresponding authors). "EgoBrain: Synergizing Minds and Eyes For Human Action Understanding", ICLR 2026.
🗺️ Contents
- 📢 News
- 📦 What's actually in the box
- 🪑 What one session looks like
- 📂 Dataset Structure
- ⏱️ How the modalities are aligned
- 🔍 Explore it interactively
- ✅ Verification
- ⚠️ Known data issues
- 🔒 Privacy
- 🚫 What's not in this release
- 🚀 Quick start
- ⚙️ Processing
- 📝 Citation
- 🌟 Acknowledgement
- 📜 License
📢 News
- [2026-10-14]: 🌐 The project page is live at lin-nie.github.io/EgoBrain — including an interactive model of the recording room: hover the desk objects, the EEG cap, the camera or the wrist bands.
- [2026-10-13]: 🎉 EgoBrain-Mini is fully released. All 40 subjects (P0001–P0040) are now available — 1,562 sub-actions, 63 hours (61 + 2 calibration), four synchronised modalities.
- [2026-10-12]: Second-pass review. An independent reviewer re-checked all 40 subjects against the published tree, segment by segment, including the bundled reports and the browse index.
- [2026-10-11]: Pre-release review across the full corpus — cross-modality timing, EEG integrity and per-sensor IMU rates re-checked for every one of the 1,562 sub-actions, not a sample.
- [2026-10-10]: Released Subject P0040 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-10]: Released Subject P0039 egocentric video + EEG + IMU + audio (Phase 2). ⚠️ One of its four IMU sensors logged at a lower rate — see Known data issues.
- [2026-10-10]: Released Subject P0038 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-10]: Released Subject P0037 egocentric video + EEG + IMU + audio (Phase 2). ⚠️ One of its four IMU sensors logged at a much lower rate — see Known data issues.
- [2026-10-10]: Released Subject P0036 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-10]: Released Subject P0035 egocentric video + EEG + IMU + audio (Phase 2). ⚠️ One of its four IMU sensors logged at a much lower rate — see Known data issues.
- [2026-10-09]: Released Subject P0034 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-09]: Released Subject P0033 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-09]: Released Subject P0032 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-09]: Released Subject P0031 egocentric video + EEG + IMU + audio (Phase 2). ⚠️ One of its four IMU sensors logged at a much lower rate — see Known data issues.
- [2026-10-06]: Released Subject P0030 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-06]: Released Subject P0029 egocentric video + EEG + IMU + audio (Phase 2). ⚠️ One of its four IMU sensors logged at a much lower rate — see Known data issues.
- [2026-10-06]: Released Subject P0028 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-06]: Released Subject P0027 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-06]: Released Subject P0026 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-06]: Released Subject P0025 egocentric video + EEG + IMU + audio (Phase 2). ⚠️ One of its four IMU sensors logged at a much lower rate — see Known data issues.
- [2026-10-06]: Released Subject P0024 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-04]: Released Subject P0023 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-04]: Released Subject P0022 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-04]: Updated P0008: 15 of its 39 sub-actions now carry EEG quality flags (a loose electrode, twice) — see Known data issues.
- [2026-10-04]: Released Subject P0021 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-04]: Released Subject P0020 egocentric video + EEG + IMU + audio (Phase 2). ⚠️ In 32 of its 39 sub-actions an EEG reference electrode had lost contact — see Known data issues.
- [2026-10-03]: Released Subject P0019 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-10-03]: Released Subject P0018 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-27]: Released Subject P0017 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-21]: Released Subject P0016 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-20]: Released Subject P0015 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-16]: Released Subject P0014 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-14]: Released Subject P0013 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-12]: Released Subject P0012 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-09]: Released Subject P0011 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-07]: Released Subject P0010 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-06]: Released Subject P0009 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-06]: Released Subject P0008 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-04]: Released Subject P0007 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-03]: Released Subject P0006 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-02]: Released Subject P0005 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-09-02]: Released Subject P0004 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-08-20]: Released Subject P0003 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-08-19]: Released Subject P0002 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-08-19]: Released Subject P0001 egocentric video + EEG + IMU + audio (Phase 2).
- [2026-08-18]: Created the Hugging Face repository for ut-vision/EgoBrain-Mini.
📦 What's actually in the box
40 subjects (P0001–P0040), 1562 real-world sub-actions (39 per subject; P0037 has 41), four modalities — first-person video, EEG, a body-worn IMU, and audio — all recorded from the same person doing the same thing at the same instant. Every one of the 1562 was verified individually, not sampled.
One sub-action from each of the 40 subjects, in order P0001–P0040 — 25 different everyday activities, four from each of the ten activity groups, every frame from that subject's own first-person camera.
| Modality | File | Format | Shape | Rate / Size | Notes |
|---|---|---|---|---|---|
| 🧠 EEG (raw) | eeg/raw.npy |
float32 array, µV | [32, T] |
32 channels @ 256 Hz | Channel-selected only, otherwise unprocessed |
| 🧠 EEG (clean) | eeg/clean.npy |
float32 array, standardized | [32, T] |
32 channels @ 200 Hz | Detrended, 0.1–75 Hz bandpass + 50 Hz notch, RANSAC bad-channel interpolation, average reference, exponential-moving standardization — see Processing |
| 🖐️ IMU | imu/imu.parquet |
Parquet, long format | [T, 20] per device |
4 body-worn sensors, rate varies by subject/session — see meta.json's imu_hz, don't assume a fixed number |
Left/right wrist + left/right upper arm — see IMU columns |
| 🎙️ Audio | audio/ego.wav |
WAV, mono PCM16 | [T] |
16 kHz | The Ego camera's own microphone |
| 🎥 Video | video/ego.mp4 |
H.264 / MP4 | — | 854×480 @ ~30 fps | First-person; a copy of its own mic audio is embedded too, but audio/ego.wav is the higher-fidelity source |
| 📄 Provenance | video/ego.json |
JSON | — | — | Exact raw source chapter + frame range this clip was cut from (for traceability; raw video itself is not distributed) |
T is the total sample count for that one sub-action's clip — duration_sec × sample_rate — not a
separate dimension on top of the rate. For example, the 76.2 s setup-calibration sub-action has
eeg/raw.npy.shape == (32, 19508) (76.2 × 256 Hz) and eeg/clean.npy.shape == (32, 15240) (76.2 × 200 Hz
exactly). Since duration varies per sub-action (from ~0.3 s to ~18 min across the 1562 of them), T is different
for every file — it's never a fixed constant you can hard-code.
The array itself carries no metadata — it's just numbers. To go the other way (shape → duration), either
read meta.json's duration_sec directly (the authoritative source, already used in Quick start),
or divide back out: duration_sec ≈ T / sample_rate (e.g. 15240 / 200 = 76.2), since the sample rate
per file is fixed and documented above.
Every one of the 1562 sub-actions has all six of these. Nothing here is a placeholder or a partial modality.
🪑 What one session looks like
(a) The desk, and the rig: a GoPro mounted on an Emotiv EEG headset, so the camera moves with the head and the electrodes stay put. (b) What that rig produces — the first-person video and the 32 EEG traces, from the same instant. (c) The participant, seated, with a monitor in front of them. (d) How one sub-action gets its boundaries: a screen prompt names the task, the participant executes it, then clicks end — and that bracket, not a human scrubbing a timeline afterwards, is what a segment in this dataset is.
| In the figure | In the files |
|---|---|
| 🧠 The Emotiv headset | 32-channel EEG at 256 Hz — eeg/raw.npy, and the processed eeg/clean.npy at 200 Hz |
| 🎥 The GoPro on it | the first-person view — video/ego.mp4, and its microphone as audio/ego.wav |
| 🖐️ Bands on both wrists and upper arms | four body-worn IMUs — imu/imu.parquet |
| 🖥️ The screen prompt | the task the participant was given — meta.json's action |
| ⏱️ Prompt → execute → end click | one sub-action's extent — meta.json's duration_sec |
1562 of these, from 40 people, with all four streams running at once.
📂 Dataset Structure
EgoBrain-Mini/
├── assets/
├── P0001/
│ ├── manifest.json ← index of all 39 segments + metadata
│ ├── segments/
│ │ └── phase2/
│ │ ├── P0001_phase2_000_setup-calibration/
│ │ │ ├── meta.json ← action label, timestamps, duration, modality flags,
│ │ │ │ and eeg_quality_flags where EEG quality is compromised
│ │ │ ├── eeg/
│ │ │ │ ├── raw.npy
│ │ │ │ └── clean.npy
│ │ │ ├── imu/
│ │ │ │ └── imu.parquet
│ │ │ ├── audio/
│ │ │ │ └── ego.wav
│ │ │ └── video/
│ │ │ ├── ego.mp4
│ │ │ ├── ego.json
│ │ │ └── redaction_windows.json ← only present where needed, see Privacy below
│ │ └── ... (39 segments total, same layout)
│ └── reports/ ← bundled, offline-first HTML report, see below
├── P0002/ ← same layout, 39 segments
├── ... ← more subjects
└── README.md
<segment_id> looks like P0001_phase2_022_drink-water-b — subject, phase, sequence number, and a
human-readable action slug.
IMU columns
imu.parquet is long-format: one row per sample, a device column (imu_1…imu_4) selecting which
sensor, and a ts column (Unix timestamp):
| Column | Meaning |
|---|---|
AccX_g, AccY_g, AccZ_g |
Acceleration X/Y/Z (g) |
GyroX_dps, GyroY_dps, GyroZ_dps |
Angular velocity X/Y/Z (degrees/second) |
AngleX_deg, AngleY_deg, AngleZ_deg |
Orientation angle X/Y/Z (degrees) |
MagX_uT, MagY_uT, MagZ_uT |
Magnetic field X/Y/Z (µT) |
QuatW, QuatX, QuatY, QuatZ |
Orientation quaternion |
Temperature_C |
Sensor temperature (°C) |
Battery_pct |
Battery level (%) |
imu_1 = left wrist, imu_2 = left upper arm, imu_3 = right wrist, imu_4 = right upper arm.
The IMU's own sample rate is not the same across subjects* — P0001 and P0022 onward logged at
~200 Hz (the intended setting), but P0002–P0021 logged at ~10 Hz for their entire sessions (a ~20x
difference; same hardware, a different logging configuration — P0002's raw column names were also
formatted differently, see above). Don't hard-code a rate: each segment's meta.json has an imu_hz
field computed directly from that file's own ts column, and that's the number to use for any
T ↔ duration conversion involving imu.parquet.
At ~200 Hz, timestamps come in batches. The sensors send samples to the logger in small Bluetooth
batches, and the logger stamps every sample of a batch with the batch's arrival time — so most rows
(80–90%) share the previous row's ts, with batches ~30–40 ms apart. The samples themselves are real
(the values change from row to row); only the per-sample time is coarse. If you need one time per
sample, spread each batch evenly up to the next batch, per device:
def respace(ts, imu_hz):
"""ts: one device's timestamps, sorted. Spread each batch evenly up to the next batch."""
ts = np.asarray(ts, float); out = ts.copy()
starts = np.flatnonzero(np.r_[True, np.diff(ts) > 0])
ends = np.r_[starts[1:], len(ts)]
for a, b in zip(starts, ends):
nxt = ts[b] if b < len(ts) else ts[a] + (b - a) / imu_hz
out[a:b] = ts[a] + (nxt - ts[a]) * np.arange(b - a) / (b - a)
return out
dev = imu[imu["device"] == "imu_1"].sort_values("ts", kind="stable")
t = respace(dev["ts"].to_numpy(), meta["imu_hz"]) # strictly increasing, ~imu_hz apart
* During early data collection (P0002–P0021), the IMU devices weren't configured to
their maximum output rate (the team was still getting familiar with the hardware) — this is why the
main EgoBrain release doesn't include IMU for those subjects at all. EgoBrain-Mini releases it anyway
rather than dropping it, since a lower-rate signal is still real, usable data — just be aware of it via
imu_hz rather than assuming parity with P0001 and P0022 onward. We're still looking into whether this can be backfilled
or otherwise compensated for in a future update.
⏱️ How the modalities are aligned
Multimodal synchronization is achieved through an EEG-anchored, cross-modal calibration procedure: EEG is treated as the ground-truth clock, and every other modality's timeline is corrected to match it — not aligned pairwise against each other. The alignment was independently re-verified for every released subject against 100% of the published sub-actions (not a sample) before release, including an EEG interval-marker cross-check baked directly into the bundled report's EEG chart (see below).
🔍 Explore it interactively
Nothing to download: the project page has an interactive model of the recording room — tap the desk objects, the EEG cap, the camera or the wrist bands to see which part of the dataset each one stands for.
Each subject's own <subject>/reports/ (e.g. P0001/reports/, P0002/reports/) is a self-contained,
offline-first HTML report — no server, no internet connection, no dependencies. Download this repo,
then just open <subject>/reports/index.html in a browser:
- A gallery of that subject's 39 sub-actions with thumbnails
- Per-action pages with synced video + live EEG / IMU / audio charts (drag the video, the charts scroll with it, and vice versa — click any chart to jump the video there)
- Plain-language explanations of what each chart actually shows and why
(HF's own file viewer won't execute the JavaScript in these pages — you do need the files on your own machine. That's by design: it means the report works identically whether you're online or not.)
Prefer to browse without downloading first? The Dataset Viewer above (the "Data Studio" /
table tab on this page) shows every sub-action across every subject as one row — thumbnail, subject,
action label, phase, duration, and which modalities it has — before you commit to downloading anything.
It's generated from browse/segments.parquet, a lightweight index (thumbnails only, no video/EEG/IMU
payload) built purely for this at-a-glance browsing; the real per-segment data still lives under each
subject's own segments/phase2/<segment_id>/ folder as described above.
✅ Verification
Every one of the 1562 sub-actions — not a sample — was individually checked:
- EEG integrity: all 32 channels finite, no flat/dead channels
- Alignment: every sub-action's report carries the EEG hardware's own interval-marker timestamp drawn on its EEG chart, so the cut can be checked against the EEG's own log segment by segment. The alignment itself was fitted per subject and then verified frame by frame at independent anchors — an on-screen task banner switching against the EEG marker, plus desk-clock ticks where a clock is legible — and reviewed by hand for every released subject. The number of anchors varies by subject (five per cross-check for P0014 onward, where the anchor list is recorded; P0001-P0013 were verified the same way under an earlier generation of the tooling)
- Coverage: 1562 / 1562, no segments skipped or excluded from verification
- EEG quality (every released subject): two checks over the whole session — the headset's own contact-quality readings for its two ear reference electrodes, and the raw amplitude (a large artifact shared across channels, sustained above ~1 mV). Affected spans are flagged per segment — see Known data issues
Check it yourself
Every subject folder ships verify_alignment.py — a standalone script that depends on nothing but
the files you downloaded (numpy, pandas + pyarrow, and ffmpeg on PATH), not on the pipeline that
produced them. It finds the same physical movements in each stream and compares when they happen.
| Stream | What it measures |
|---|---|
| 🎥 Video | frame-to-frame change in the lower two thirds of the ego view, where the hands are |
| 🖐️ IMU | angular speed, all four sensors together and each one on its own |
| 🧠 EEG | movement artifacts — 1–15 Hz activity in eeg/raw.npy |
Each curve is slid against the video's; the shift with the best match is that segment's lag. The verdict uses the subject's median lag, since a single segment scatters by a few tenths of a second.
python P0001/verify_alignment.py P0001
Run over all 40 subjects on the released files, every subject passes on every stream:
| Stream vs ego video | Subjects passing | Median across subjects | Largest subject median |
|---|---|---|---|
| 🧠 EEG | 40 / 40 | −0.00 s | 0.19 s |
| 🖐️ IMU, four sensors together | 40 / 40 | −0.02 s | 0.05 s |
imu_1 left wrist |
40 / 40 | −0.02 s | 0.06 s |
imu_2 left upper arm |
40 / 40 | −0.00 s | 0.06 s |
imu_3 right wrist |
40 / 40 | −0.04 s | 0.07 s |
imu_4 right upper arm |
40 / 40 | −0.02 s | 0.14 s |
The motion check only looks at sub-actions longer than 45 s (1260 of the 1562) — shorter ones rarely hold enough movement for a confident match. Of those it found a clear match in 1255 for the IMU and 988 for EEG, whose movement artifact is much weaker than a gyroscope's signal. Segments without a clear match are left out of the median rather than counted as zero.
⚠️ Known data issues
Known problems in the released data. The files themselves are left unaltered; every affected range is
listed in that segment's meta.json (eeg_quality_flags) and marked in the bundled report.
| Subject | Modality | Issue | Affected range | Segments | Recommendation |
|---|---|---|---|---|---|
| P0008 | EEG | An electrode came loose twice; the signal is dominated by a 1–3 mV artifact shared across channels (normal EEG: ~0.05–0.15 mV) | phase2_021 1 s → phase2_025 249 s, and phase2_027 1 s → phase2_036 185 s |
15 of 39 | Exclude from EEG analyses |
| P0020 | EEG | An ear reference electrode (CMS/DRL) lost contact; all 32 channels are affected (alpha power ~85% lower; blinks and chewing still visible) | phase2_007 111.3 s → phase2_038 36.4 s (contact briefly back in phase2_019, 154.5–217.5 s) |
32 of 39 | Exclude from band-power / amplitude-based analyses |
| P0025 | IMU | One of the four sensors (imu_3, right wrist) logged at ~58 Hz while the other three held ~198.6 Hz — Bluetooth packet loss on that unit, not a configuration difference |
The whole session | 39 of 39 | Use meta.json's imu_hz_per_device; treat imu_3 as sparse, and avoid analyses that assume the four sensors share a sample grid |
| P0029 | IMU | One of the four sensors (imu_2, left upper arm) logged at ~62.8 Hz while the other three held ~198.6 Hz — the same Bluetooth packet loss as P0025, on a different unit |
The whole session | 39 of 39 | Use meta.json's imu_hz_per_device; treat imu_2 as sparse, and avoid analyses that assume the four sensors share a sample grid |
| P0031 | IMU | One of the four sensors (imu_1, left wrist) logged at ~61 Hz while the other three held ~198.6 Hz — the same Bluetooth packet loss as P0025 and P0029, on a different unit again |
The whole session | 39 of 39 | Use meta.json's imu_hz_per_device; treat imu_1 as sparse, and avoid analyses that assume the four sensors share a sample grid |
| P0035 | IMU | One of the four sensors (imu_4, right upper arm) logged at ~159 Hz while the other three held ~198.6 Hz — the same Bluetooth packet loss as P0025, P0029 and P0031, on a different unit again |
The whole session | 39 of 39 | Use meta.json's imu_hz_per_device; treat imu_4 as sparse, and avoid analyses that assume the four sensors share a sample grid |
| P0037 | IMU | One of the four sensors (imu_1, left wrist) logged at ~49 Hz while the other three held ~198.6 Hz — the same Bluetooth packet loss as P0025, P0029, P0031 and P0035, and the slowest case in the corpus |
The whole session | 41 of 41 | Use meta.json's imu_hz_per_device; treat imu_1 as sparse, and avoid analyses that assume the four sensors share a sample grid |
| P0039 | IMU | One of the four sensors (imu_2, left upper arm) logged at ~152 Hz (range 137–175) while the other three held ~196–199 Hz — the mildest case of the same Bluetooth packet loss as P0025, P0029, P0031, P0035 and P0037; it is listed because the gap is still large enough to break a shared sample grid |
The whole session | 39 of 39 | Use meta.json's imu_hz_per_device; treat imu_2 as sparse, and avoid analyses that assume the four sensors share a sample grid |
For the EEG rows, video, audio and IMU are unaffected; for the IMU rows, video, audio and EEG are unaffected.
How they were found. The two EEG cases were caught by the headset's own contact-quality
channels together with the raw amplitude: in both, an experimenter is seen re-attaching an
electrode, and the EEG shows a sharp transient at the loss and at the re-attachment. The IMU
cases were caught by a per-subject rate check run before every release — meta.json's imu_hz
averages the four sensors and so hides a single slow unit, while imu_hz_per_device exposes it:
"imu_hz": 161.38,
"imu_hz_per_device": {"imu_1": 198.67, "imu_2": 198.67, "imu_3": 49.51, "imu_4": 198.67}
The samples a slow unit did record are real and correctly timestamped — there are simply fewer of them. Each affected subject's bundled report carries a notice above its IMU chart, so the sparser trace is explained where it is seen.
Using the flags. Segments without eeg_quality_flags have no known EEG issue. rel_start /
rel_end are seconds from the segment start; issue is signal_quality_poor or
reference_contact_lost:
"eeg_quality_flags": [
{"rel_start": 0.0, "rel_end": 154.5, "issue": "reference_contact_lost",
"description": "...", "evidence": "..."}
]
To keep only unflagged EEG samples:
flags = meta.get("eeg_quality_flags", [])
fs = 200 # clean.npy rate
keep = np.ones(eeg_clean.shape[1], bool)
for f in flags:
keep[int(f["rel_start"] * fs):int(f["rel_end"] * fs)] = False
eeg_ok = eeg_clean[:, keep] # samples outside every flagged range
🔒 Privacy
131 of the 1562 sub-actions briefly show something that needed redacting — most often a laptop's own
lock screen or an app window, a hand-held phone, or earlier pages of a shared notebook being leafed
through; occasionally a bystander incidentally in frame. Rather
than blank out the whole picture for something covering a fraction of it, redaction is region-limited
blur: only the specific area of the frame containing the sensitive detail is blurred (strong enough
that text, faces, and names are unreadable), for only the seconds it's actually visible — the rest of
the frame, and the rest of the clip, is untouched. A small number of windows also mute audio, when the
audio itself (not just the picture) needed it. This is baked into the video at the source, not a
runtime toggle. video/redaction_windows.json documents exactly which seconds and which region of the
frame were affected, for transparency. Nothing else in the release is redacted.
If you come across any content in this release that you believe may reveal a subject's private information, please contact the maintainers immediately at nielin@iis.u-tokyo.ac.jp and it will be re-processed and re-uploaded.
🚫 What's not in this release
- Raw, unprocessed video — not distributed at all;
video/*.jsonretains exact provenance (source file + frame range) in case higher-fidelity access is ever needed for approved use. - Only Phase 2 (real-world action) sub-actions are included. Phases 1 and 3 (the in-lab calibration and resting blocks) stay in the full EgoBrain release.
🚀 Quick start
import json
import numpy as np
import pandas as pd
root = "P0001/segments/phase2/P0001_phase2_022_drink-water-b"
meta = json.load(open(f"{root}/meta.json"))
eeg_clean = np.load(f"{root}/eeg/clean.npy") # [32, T] @ 200 Hz
imu = pd.read_parquet(f"{root}/imu/imu.parquet") # long format, 4 devices
print(meta["action"], meta["duration_sec"], "sec")
print("EEG:", eeg_clean.shape)
print("IMU devices:", imu["device"].unique())
⚙️ Processing
eeg/clean.npy is produced by (in order): linear + 6th-order detrending → 0.1–75 Hz bandpass filter →
50 Hz notch filter → resample to 200 Hz → RANSAC-based bad-channel detection and interpolation → average
reference → exponential-moving-average standardization. eeg/raw.npy skips all of this — it's the
channel-selected signal at its native 256 Hz, in µV.
📝 Citation
If EgoBrain-Mini is useful in your research, please cite the paper:
@inproceedings{lin2026egobrain,
title = {EgoBrain: Synergizing Minds and Eyes For Human Action Understanding},
author = {Nie Lin and Yansen Wang and Dongqi Han and Weibang Jiang and
Jingyuan Li and Ryosuke Furuta and Yoichi Sato and Dongsheng Li},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2026},
url = {https://arxiv.org/abs/2506.01353}
}
Please also mention the release you used, so results stay reproducible:
EgoBrain-Mini (Hugging Face:
ut-vision/EgoBrain-Mini), 40 subjects, 1,562 sub-actions.
🌟 Acknowledgement
We sincerely thank all participants who contributed their time and effort to the data collection process.
This project was initiated and developed during Nie Lin's research internship at Microsoft Research Asia (MSRA), and later continued at the Institute of Industrial Science (IIS), The University of Tokyo. We thank all collaborators and mentors at MSRA and IIS for their valuable guidance and support.
We also acknowledge the support from our advisors, collaborators, and the community.
📜 License
EgoBrain-Mini is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.
By using this dataset, you agree to:
- Use the dataset only for non-commercial research purposes
- Not redistribute the data
- Not attempt to identify the participant
- Comply with applicable ethical and data protection regulations
If you have questions regarding licensing or usage, please contact the maintainers at nielin@iis.u-tokyo.ac.jp.
License details: https://creativecommons.org/licenses/by-nc/4.0/
- Downloads last month
- 2,482