humyn-labs/Egocentric-3-Camera-Array

Egocentric 3-Camera Array (Head + Both Wrists)

Three long-form household and warehouse tasks recorded simultaneously from three body-mounted cameras — head, left wrist and right wrist — each with per-frame timestamps and its own high-rate gyroscope and accelerometer.

This is a bimanual manipulation dataset: the wrist cameras see what each hand is doing at close range while the head camera carries the scene context.

Preview: 45 s of the Cooking task, all three views at the same instant. Left panel = head, middle = left wrist, right = right wrist. Built for this card only; the repo ships each view as a separate file.

---

At a glance

Tasks	3 (Sort & Repack · Cooking · Packing T-shirts)
Cameras per task	3 (head, left wrist, right wrist)
Activity duration	60.9 min
Total video	182.6 min across 9 files
Resolution	1440×1080 (4:3) @ 29.97 fps
Audio	none
Video frames	328,129 (with per-frame timestamps)
Gyroscope samples	17,522,016 (~1576–1621 Hz per device)
Accelerometer samples	2,189,132 (~197–203 Hz per device)
Files	36 = 3 tasks × 3 cameras × (video + times + gyro + accel)

Task	Duration	Frames (head/L/R)
Sort & Repack	20.11 min	36141 / 36142 / 36142
Cooking	20.07 min	36075 / 36075 / 36075
Packing T-shirts	20.68 min	37159 / 37160 / 37160

---

Synchronisation — read this first

Each camera keeps its own clock, and every clock starts at 0. There is no shared sync signal, no common epoch, and no clapperboard event in the released files. Alignment is by the assumption that all three devices started together.

That assumption holds well but not perfectly. Measured drift between the three clocks at end of recording:

Task	Spread across the 3 cameras	In frames @ 29.97 fps
Sort & Repack	33.4 ms	~1.0
Cooking	19.1 ms	~0.6
Packing T-shirts	32.8 ms	~1.0

So cross-camera alignment is good to about one frame over a 20-minute recording, and frame counts differ by at most 1 between cameras. That is fine for action recognition and coarse fusion; it is not good enough for anything needing sub-millisecond stereo-grade sync. Each device also runs a slightly different IMU rate (1576 vs 1604 vs 1621 Hz), which is the same independent-oscillator effect.

Within a single camera, video and IMU do share a timebase — video_times.csv, gyro.csv and accel.csv all use the same t_seconds column, so per-camera fusion is exact.

import csv, bisect

times = [float(r["t_seconds"]) for r in csv.DictReader(open("timestamps/cooking_head_video_times.csv"))]
gyro  = [(float(r["t_seconds"]), float(r["gx"]), float(r["gy"]), float(r["gz"]))
         for r in csv.DictReader(open("imu/cooking_head_gyro.csv"))]

def gyro_at_frame(i):                      # nearest gyro sample to frame i
    t = times[i]
    k = bisect.bisect_left(gyro, (t,))
    return min(gyro[max(0, k-1):k+1], key=lambda g: abs(g[0] - t))

---

Repository layout

data/train-*.parquet   # 720p previews of all three views + metadata (powers the viewer)
videos/*.mp4           # full-resolution 1440×1080 captures, 9 files
timestamps/*_video_times.csv   # frame_idx, t_seconds
imu/*_gyro.csv                 # t_seconds, gx, gy, gz
imu/*_accel.csv                # t_seconds, ax, ay, az
preview/               # 720p proxies; card_sample_3cam_montage.mp4 is the 3-up clip above
metadata.csv           # flat table

One row per task, not per camera — each row carries all three views and all nine sidecar files.

Columns

Column	Description
video	720p preview of the head camera — plays in the viewer
left_wrist_preview_video, right_wrist_preview_video	720p previews of the wrist cameras
sample_id, task	e.g. cooking / Cooking
{cam}_video_path	Full-resolution file, for cam in head, left_wrist, right_wrist
{cam}_times_path, {cam}_gyro_path, {cam}_accel_path	Sidecar CSVs
{cam}_duration_seconds, {cam}_width, {cam}_height, {cam}_fps	Probed from the media
{cam}_frames	Rows in that camera's video_times.csv
{cam}_gyro_rows, {cam}_accel_rows	IMU sample counts
{cam}_gyro_hz, {cam}_accel_hz	Measured rate, not nominal
{cam}_gyro_columns, {cam}_accel_columns	CSV headers
{cam}_times_t_end, {cam}_gyro_t_end, {cam}_accel_t_end	Last timestamp — use these to check drift
{cam}_bytes, {cam}_sha256	Size and integrity of the original

Units

Accelerometer is in m/s² and gyroscope in rad/s, both verified empirically rather than assumed: resting accelerometer magnitude has a median of 9.83 across samples, and gyroscope magnitude sits at 3.5 rad/s (200 °/s) at the 99th percentile, which is the expected range for wrist motion. Note this differs from the LATAM residential release, whose accelerometer is in g — do not mix the two without rescaling.

---

Usage

from datasets import load_dataset

ds = load_dataset("humyn-labs/Egocentric-3-Camera-Array", split="train")
r = ds[0]
print(r["task"], r["head_frames"], r["head_gyro_hz"], "Hz")

Full-resolution video and all IMU:

from huggingface_hub import snapshot_download
snapshot_download("humyn-labs/Egocentric-3-Camera-Array", repo_type="dataset",
                  allow_patterns=["videos/*", "imu/*", "timestamps/*"])

Intended uses

Bimanual manipulation · multi-view action recognition · hand-activity classification from wrist cameras · video + IMU sensor fusion · viewpoint-invariant representation learning · long-horizon procedural task segmentation · imitation learning for two-armed robots.

Limitations

• Three recordings. Long (~20 min each) but only three tasks, with no held-out split. The source data carries no subject identifiers, so subject diversity cannot be established from this release.
• Cross-camera sync is implicit and drifts ~1 frame, as described above.
• No action labels, no annotations, no captions. This release is raw sensor data only.
• No magnetometer and no camera calibration, so no absolute orientation and no metric 3D.
• 4:3 aspect at 1440×1080, unlike the 16:9 clips elsewhere in this collection — check your resize path.
• Wide-angle lenses produce noticeable barrel distortion, uncorrected and with no distortion coefficients supplied.
• No audio.

Provenance

Curated from the HumynLabs egocentric sample collection. All technical fields — durations, frame counts, IMU rates and clock spans — were measured from the files rather than copied from the source sheet. All 36 files resolved and downloaded with matching byte sizes; nothing was dropped.

License

CC BY 4.0. Recorded with participant consent for research use.

Privacy note specific to this release: the wrist-mounted cameras point back toward the wearer for much of each recording and frequently capture the wearer's face, which head-mounted egocentric footage does not. Home and workplace interiors and incidental bystanders also appear. Please handle accordingly and do not attempt to identify individuals.

Citation

@misc{humynlabs2026egocentric3cam,
  title  = {Egocentric 3-Camera Array (Head + Both Wrists)},
  author = {HumynLabs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/humyn-labs/Egocentric-3-Camera-Array}
}