A privacy-redacted, time-aligned, multi-camera + IMU egocentric dataset captured with the Human Archive GSI Cap — a head-worn 6-camera rig with onboard IMUs designed for dexterous-manipulation and long-horizon-task research.

Every clip in this release has passed:

automated quality screening (Gemini-based QA),
frame-accurate cutting to within ≤ 1 frame across all six camera streams,
EgoBlur Gen 2 face redaction on every frame,
ffprobe verification for codec / duration / frame count / cross-camera consistency,
task / environment / scene labelling via Gemini 2.5 Pro on the native video stream.

Total at release: 484 clips · 2,904 camera mp4s · ~23.9 hours of footage · 22 semantic groups.

1. Directory layout

Clips are organised by semantic group rather than by recording session. Each group bucket holds clips with conceptually similar tasks and/or environments, and each clip's leaf folder is fully self-contained — every file you need to use that clip (videos, IMU, calibration, metadata) lives inside it.

HA-Ego-1K/
├── README.md
├── <semantic-group>/                           <-- 24 groups; see §2
│   └── <description_slug>_<N>/                 <-- 484 leaf folders total; N draws from 1..519 (hospital clips reserved for an internal release)
│       ├── video_dev0_..._right-eye_clipNN.mp4
│       ├── video_dev1_..._left-front_clipNN.mp4
│       ├── video_dev2_..._left-eye_clipNN.mp4
│       ├── video_dev3_..._right_clipNN.mp4
│       ├── video_dev4_..._left_clipNN.mp4
│       ├── video_dev5_..._right-front_clipNN.mp4
│       ├── imu.csv                             <-- clip-aligned (t=0 at clip start)
│       ├── metadata.json                       <-- task / env / source_id / clip_n / device_uid
│       └── calibration/                        <-- this device's Kalibr outputs
│           ├── imus_intrinsic/
│           ├── imus_cam_l_extrinsic/
│           ├── imus_cam_r_extrinsic/
│           ├── imus_cam_lr_front_extrinsic/
│           └── imus_cam_lr_eye_extrinsic/

Leaf folder name

Each leaf folder is <description_slug>_<N> — for example, car_door_audio_system_installation_142/.

<description_slug> is a slugified form of the clip's high_level_description (lowercased, non-alphanumeric → _, capped at 60 chars).
<N> is a globally unique integer assigned in deterministic order across the entire dataset. The 484 public clips use N drawn from 1..519; the 35 N values not present here are hospital clips kept in an internal release.

metadata.json

Every leaf folder includes a metadata.json with the full provenance and labels for that clip:

{
  "n": 142,
  "source_id": "01KSD9R1QZN8RSQ44FXSYNCHC3",
  "clip_n": 3,
  "device_uid": "6475c570c8c17f0f",
  "task": "Mechanical repair",
  "environment": "Car Shop",
  "high_level_description": "car door audio system installation",
  "folder_name": "car_door_audio_system_installation_142"
}

n — global clip index (1..519).
source_id — 26-char ULID of the original recording session (one cap, one continuous capture). Multiple clips can share a source_id.
clip_n — index within the source recording (kept-segment number; 1, 2, … in composite time order).
device_uid — 16-char hex; identifies which cap recorded the clip. Used to pick the right calibration — but note calibration is also copied into the leaf folder itself, so most users won't need to dereference this.
task / environment — short labels from the QA pass.
high_level_description — free-form one-line description (the basis of the folder slug and the semantic grouping).

2. The 22 semantic groups

Clips are grouped by semantic similarity of high_level_description so that related activities live next to each other on disk.

group	# clips	hours	examples
`car_workshop`	48	1.66	car audio install, interior detailing, electrical wiring
`construction_tile_panel_tool`	39	1.32	tile cutting, panel install, power-tool use
`bottle_factory`	36	0.98	bottle moulding, conveyor loading, QA inspection
`warehouse_packing`	32	1.85	packing boxes, applying shipping labels, sorting
`home_cleaning`	30	2.00	dusting, mopping, sweeping, surface wiping
`laundry_sorting_folding`	30	1.02	sorting / folding / hanging clothes
`shoe_washing`	30	1.44	hand-washing & cleaning shoes
`laundry_handwashing`	29	0.85	hand-washing garments in sink / basin
`factory_misc`	22	0.63	non-bottle factory work — mixers, panels, fibres, embroidery
`construction_misc`	22	0.55	site upkeep, cement, plumbing pipes
`nail_salon`	21	1.97	manicures, acrylic nails, nail filing
`laundry_machines`	20	0.42	washing-machine load/unload, laundromat ops
`bar_restaurant`	20	0.90	bartending, restaurant floor cleaning
`office_electronics_repair`	16	0.61	electronics repair, electrical wiring, lab/desk admin
`construction_plastering`	15	0.42	plastering walls and ceilings
`hair_salon`	13	2.36	hair washing, cutting, styling
`nursery_plants`	13	0.28	plant care, watering, weeding, harvesting
`bathroom_cleaning`	12	0.45	bathroom cleaning, toilet, sink, mirror
`laundry_other`	11	0.33	misc laundry (stain treating, doing laundry)
`laundry_ironing`	11	2.78	ironing shirts, pants, clothes
`appliance_repair`	8	0.30	small-appliance / fan / AC repair
`kitchen_cooking_dishes`	6	0.79	food prep, cooking, dish washing
total	484	23.93

Groups were derived with rule-based clustering on task, environment, and high_level_description. They are intended as a convenience grouping, not a formal taxonomy — if you need precise filtering, read task / environment / high_level_description from each metadata.json.

3. Camera layout

Each leaf folder contains six time-aligned camera streams. Naming convention is video_dev{0..5}_session{N}_segment{M}_{camera-label}_clip{NN}.mp4. The camera-label field is the canonical name:

dev	label	physical position
0	`right-eye`	right eye-line, forward
1	`left-front`	left of nose-bridge, forward
2	`left-eye`	left eye-line, forward
3	`right`	right temple, side-facing
4	`left`	left temple, side-facing
5	`right-front`	right of nose-bridge, forward

The dev index is just the camera-tree port and is not meaningful for downstream use; prefer the label.

4. Video format & quality

property	value
container	MP4 (`isom`/`isom2`/`mp41` brands)
codec	HEVC (`hvc1` tag, NVENC-encoded on-device)
pixel format	yuv420p
resolution	1920 × 1080
frame rate	30 fps, constant
audio	none
start position	`+faststart` (moov atom at file head)

Cross-camera frame counts within a single leaf folder match within ≤ 1 frame (~33 ms drift max). This is enforced by the QA pipeline's verify_and_trim_all.py step; any clip where it could not enforce that bound was dropped from this release.

5. Face redaction (EgoBlur)

Every video frame has been processed by EgoBlur Gen 2 (face model).

Model: ego_blur_face_gen2.jit (Meta, Apache 2.0)
Detection threshold: 0.3 (tuned for the cap's 175° fisheye + libx264 cut artifacts; lower than Aria's 0.674 because the wider-FOV optics make faces smaller and detection scores lower).
Coverage: every frame of every camera; the blur is rasterised into the file itself, there is no separate mask track.

You do not need to do any blurring on your end. Blurred regions are elliptical Gaussian masks rasterised into the pixel data.

6. IMU data (`imu.csv`)

Each leaf folder includes a single imu.csv synthesising the GSI Cap's onboard IMUs over the time span of that clip.

Schema

column	unit	source
`t_sec`	seconds	re-zeroed so that `t_sec = 0` is the first video frame of the clip
`ax`, `ay`, `az`	m/s²	linear acceleration
`gx`, `gy`, `gz`	°/s	angular velocity
`mx`, `my`, `mz`	µT	magnetometer (if available; NaN where missing)

Sensor specs

Accelerometer: ICM-42688-P, range ± 4 g (≈ ± 39.2 m/s²), 16-bit raw, scale factor 0.001197 m/s²/LSB.
Gyroscope: ICM-42688-P, range ± 500 °/s, 16-bit raw, scale factor 0.01526 °/s/LSB.
Magnetometer: MMC5983MA, scale 0.00625 µT/LSB, ~110 Hz target.

Effective sample rate ranges from ~12 Hz to ~22 Hz in practice (target was 50 Hz; BLE-link bandwidth is the bottleneck). Down-sampling is left to the consumer.

To align IMU samples to a particular video frame, look up the IMU row whose t_sec is closest to frame_index / 30.0.

7. Calibrations (`calibration/`)

Every multi-cam device used in this dataset has been Kalibr-calibrated for both the camera intrinsics and the IMU-to-camera extrinsics.

In the new layout, calibration is copied into every leaf folder — there is no shared top-level calibrations/ directory. Two leaf folders recorded with the same cap will contain byte-identical calibration files. This keeps every clip self-contained at the small cost of duplication.

Each calibration/ subfolder contains:

subfolder	what it calibrates
`imus_intrinsic/`	IMU intrinsic params (bias, noise, random walk)
`imus_cam_l_extrinsic/`	IMU ↔ left side camera (`video_dev4`)
`imus_cam_r_extrinsic/`	IMU ↔ right side camera (`video_dev3`)
`imus_cam_lr_front_extrinsic/`	IMU ↔ left-front + right-front (`video_dev1`, `video_dev5`)
`imus_cam_lr_eye_extrinsic/`	IMU ↔ left-eye + right-eye (`video_dev2`, `video_dev0`)

Each subfolder follows Kalibr's standard output convention:

*-camchain.yaml — camera intrinsics (focal length, principal point, distortion model, image resolution) per camera in the chain.
*-camchain-imucam.yaml — IMU-to-camera transforms (T_cam_imu) per camera in the chain.
*-imu.yaml — IMU noise parameters.
*-results-*.txt — human-readable Kalibr report.

To dedupe across clips at consumer time, key on metadata.json → device_uid; all leaf folders sharing a device_uid have identical calibration/ contents.

8. Quick start — load one clip end-to-end

import pathlib, json, pandas as pd, cv2, yaml

root = pathlib.Path("HA-Ego-1K")
clip = root / "car_workshop" / "car_door_audio_system_installation_142"

# Metadata + labels
meta = json.load(open(clip / "metadata.json"))
print(meta["task"], "|", meta["environment"], "|", meta["high_level_description"])

# IMU (already clip-aligned: t_sec = 0 is the first video frame)
imu = pd.read_csv(clip / "imu.csv")
print(imu.head())

# Video (any of the six cameras)
left_front = next(clip.glob("video_dev1_*_left-front_clip*.mp4"))
cap = cv2.VideoCapture(str(left_front))
ok, frame = cap.read()
print(frame.shape)              # (1080, 1920, 3)

# Calibration (self-contained inside the leaf folder)
camchain = yaml.safe_load(open(clip / "calibration"
                                    / "imus_cam_lr_front_extrinsic"
                                    / "cam_lr_front_intrinsic-camchain.yaml"))
print(camchain["cam0"]["intrinsics"])         # [fx, fy, cx, cy]
print(camchain["cam0"]["distortion_coeffs"])
print(camchain["cam0"]["distortion_model"])   # "equidistant" / "radtan" / ...

Filter by task or environment

Because every clip has a metadata.json, you can ignore the group-folder organisation and filter purely on labels:

import json, pathlib

root = pathlib.Path("HA-Ego-1K")
all_meta = [json.load(open(p)) for p in root.rglob("metadata.json")]

# Every laundry-room clip:
laundry = [m for m in all_meta if m["environment"] == "Laundry Room"]
print(len(laundry))

# Every clip with description containing "ironing":
ironing = [m for m in all_meta if "ironing" in m["high_level_description"]]
print(len(ironing))

9. Cross-camera & IMU alignment

All six camera streams in a leaf folder share the same composite_start / composite_end boundary from the original recording's manifest. Concretely:

Pick a frame index i (0-based) in any of the six MP4s.
The corresponding wall-clock time relative to clip start is t = i / 30.0 seconds.
The IMU row whose t_sec is closest to t is the aligned IMU reading.
The same t indexes the same physical moment in all other five cameras.

If you ever observe more than one-frame drift (~33 ms) between cameras in a clip, it is a bug — please open an issue with the clip path.

10. Roadmap

v1.x — depth (from the OAK side cameras) on a subset.
v1.x — tactile-glove streams on a subset (left + right hand, 256 taxels each).

11. Licence & citation

Licence: CC-BY-NC-4.0 (non-commercial; attribution required).
Citation: TBD — please cite this dataset page until a paper is posted.

If you build on this dataset, please send a heads-up to team@humanarchive.ai — we'd love to know what you're doing with it.

Downloads last month: 113

Total file size:

548 GB

You need to agree to share your contact information to access this dataset

HA-Ego-1K