LATESTEgocentric-10K coverage release, warehouse + retail →
FIRSTHANDGet sample data

Sample data outputs

What you actually receive, per recording.

Every recording is a first-person video of a real task — cooking, cleaning, repair, assembly, gardening, retail — filmed head-mounted with the hands in view, at the original speed. It ships as the video plus the camera's own sensor data, extracted and cleaned up. Below are two real 10-second slices, starting with a US home-baking recording, each with the exact IMU rows that go with it.
Camera
Insta360 X5, head-mounted
Native video
4608 × 4608 · 30 fps
IMU rate
~1,000 samples / s
Files per recording
Up to 8

Sample 01 · USA

Ten seconds of a home-baking recording, filmed in the US.

Faces blurred and downscaled to 1080 × 1080 so it loads on a web page. Delivered files are the full 4608 × 4608 fisheye frame at the original length.

egocentric_baking_10s_usa.mp4 · head-mounted single-lens fisheye · hands in frame · faces blurred

Download the sample

egocentric_baking_10s_usa.mp4 ↓10 seconds of a home-baking recording. Faces blurred, downscaled to 1080 × 1080 for the web; delivered files are full 4608 × 4608.imu_sample_10s_usa.csv ↓The real IMU rows for those same 10 seconds (~10,000 samples), exactly as delivered.

The two files cover the same 10 seconds, so you can check video–IMU alignment yourself before any commercial conversation.

The real IMU rows for that same clip.

Plotted straight from imu_sample_10s_usa.csv — 9,934 rows over 10.0 seconds, drawn every 20th sample. Values are raw sensor readings; the IMU orientation, units and scale ship in camera_dump.txt.

Gyroscope · 3 axesgyroADC[0]gyroADC[1]gyroADC[2]
0s1s2s3s4s5s6s7s8s9s
Accelerometer · 3 axesaccSmooth[0]accSmooth[1]accSmooth[2]
0s1s2s3s4s5s6s7s8s9s
imu_sample_10s_usa.csv · first 5 rows
loopIteration,time,gyroADC[0],gyroADC[1],gyroADC[2],accSmooth[0],accSmooth[1],accSmooth[2]
178982,180000393,-13.203640381071500,16.048975991247000,1.032528575125110,5173.8644125071300,-19055.81884761240,1551.9628031068800
178983,180001393,-13.507031917221700,16.109170244171800,0.8462079156333540,5193.659226604000,-19036.265585441400,1532.4722955707700
178984,180002393,-13.749860174711100,15.925482249358200,0.7224073640133620,5154.616999947820,-19016.746212394900,1512.4027207138800
178985,180003393,-13.870056227018800,15.802677730951100,0.538554449045206,5095.951831345800,-18977.624968217600,1492.0688139722300
178986,180004393,-14.17281399673540,15.801615665423000,0.2914239023093320,5095.941236317630,-18958.011800712100,1491.99750452773
loopIteration
Sample counter, increments once per IMU read
time
Timestamp in microseconds, on the same timeline as the video
gyroADC[0..2]
3-axis gyroscope (x, y, z)
accSmooth[0..2]
3-axis accelerometer (x, y, z)

Standard Gyroflow / Blackbox CSV layout. A full 10-minute recording is about 600,000 rows.

Sample 02

Ten seconds of a home-cooking recording.

Faces blurred and downscaled to 1080 × 1080 so it loads on a web page. Delivered files are the full 4608 × 4608 fisheye frame at the original length.

egocentric_cooking_10s.mp4 · head-mounted single-lens fisheye · hands in frame · faces blurred

Download the sample

egocentric_cooking_10s.mp4 ↓10 seconds of a home-cooking recording. Faces blurred, downscaled to 1080 × 1080 for the web; delivered files are full 4608 × 4608.imu_sample_10s.csv ↓The real IMU rows for those same 10 seconds (~10,000 samples), exactly as delivered.

The two files cover the same 10 seconds, so you can check video–IMU alignment yourself before any commercial conversation.

The real IMU rows for that same clip.

Plotted straight from imu_sample_10s.csv — 10,053 rows over 10.0 seconds, drawn every 20th sample. Values are raw sensor readings; the IMU orientation, units and scale ship in camera_dump.txt.

Gyroscope · 3 axesgyroADC[0]gyroADC[1]gyroADC[2]
0s1s2s3s4s5s6s7s8s9s
Accelerometer · 3 axesaccSmooth[0]accSmooth[1]accSmooth[2]
0s1s2s3s4s5s6s7s8s9s

More samples

Preview other modalities and formats.

Everything below is public, so a lab can review it before any commercial conversation.

Head-mounted stereo camera rig on a worker assembling a wooden joint, with side-by-side left and right views and a depth map on the monitor
Stereo sampleA stereoscopic recordingCalibrated left + right views, per-session calibration, hardware sync and 3D labels.

Per recording

Eight outputs, one folder.

Each recording is delivered as a folder containing the files below. GPS and the camera-position photo are included when available for that recording.

  1. 01Video

    video.mp4
    • Insta360 X5, head-mounted, single-lens fisheye: first-person view with hands in frame
    • Native 4608 × 4608 at 30 fps, original length (typically about 10 min, continuous)
    • Faces detected on every frame, delivered blurred or as per-frame face bounding boxes

  2. 02IMU motion data

    imu.csv
    • 3-axis gyroscope and 3-axis accelerometer at about 1,000 samples per second
    • About 600,000 rows per 10-minute recording
    • Timestamped in microseconds and synced to the video
    • Standard Gyroflow / Blackbox CSV format

  3. 03Camera calibration + full telemetry

    camera_dump.txt
    • Per-lens fisheye calibration: camera matrix (fx, fy, cx, cy) and distortion coefficients, ready for undistortion and 3D work
    • Per-frame exposure / shutter data
    • IMU orientation, units and scale
    • Camera model, serial number, firmware, capture settings

  4. 04Camera summary

    camera_summary.json
    • Recording metadata: model, firmware, resolution, frame rate, duration, capture settings

  5. 05GPS track

    gps.csv
    • Timestamped latitude, longitude, altitude and speed, when a GPS source is paired to the camera

  6. 06Face-privacy data

    blur_face_frames.json
    • Per-frame face bounding boxes for every detected face

  7. 07Task + object labels

    object-checklist.json
    • Activity/routine name and category
    • Which target objects appear in the recording (e.g. can opener, screwdriver, cup)

  8. 08Camera-position photo

    when provided
    • A still showing how and where the camera was worn for that recording

Collection + QA

How a recording gets to your folder.

The process behind every file above. For the wider capture setup, see capture and the example footage.

01

Contributor network

Distributed contributor network with a web portal for assignment, upload and review

02

Human review

Every recording is human-reviewed before delivery: task match, hands visible, no staged or idle footage

03

Automated checks

Uploads missing camera telemetry are rejected automatically

04

File organisation

Human-readable, PII-free layout: project / category / activity / slot / version

Face privacy is handled per customer: the same per-frame detections are delivered either burned in as blur or as bounding boxes in blur_face_frames.json.

Evaluating before you commit? See how to request an egocentric data sample.

Want this for your task?

Send the activity and object list you need. Every recording comes back in the same folder structure shown on this page.

Request a custom dataset →See the full sample pack