Factory Stereo Egocentric (Sample)

Head-mounted stereo recordings of people doing their own jobs on real factory lines: synchronized stereo video, IMU, per-session calibration and 3D hand keypoints. Released for non-commercial research under CC BY-NC 4.0. This is a small preview (v0.1); a larger release is in preparation.

Access on Hugging Face is granted automatically after a short form (institution, intended use, non-commercial agreement).

Get the data on Hugging Face Data format Quality checks Cite

18episodes, about 8 minutes in total

12tasks

9sites, anonymized as site_NN

2 × 1080pstereo fisheye pair, ~30 fps

62%of frames with at least one 3D hand

Tasks

Task	Type of work
Sock knitting machine: yarn change and fault handling	Machine tending
Sock shaping and packing	Packing
Sock bundle packing	Packing
Sock packing	Packing
Toy packaging	Packing
Umbrella packing	Packing
Chopstick packaging	Packing
Labelling packaging bags	Labelling
Packaging-bag inspection	Inspection
Plastic bag making	Fabrication
Applying glue to paper	Fabrication
Metal chain-belt making	Fabrication

What each episode contains

cam0.mp4, cam1.mp4: stereo pair, frame i of one pairs with frame i of the other; faces blurred, no audio.
timestamps.json: per-frame sensor timestamps for both cameras.
imu.csv: accelerometer and gyroscope, each sample tagged with its cam0 frame.
calibration.json: fisheye intrinsics (Kannala–Brandt), extrinsics, baseline.
hands_3d.jsonl: per frame and hand, 2D keypoints and 3D keypoints in the camera frame, a wrist-centred frame and a gravity-aligned frame, with reprojection error.
lineage.json: processing steps and pipeline version.

Field-level documentation: Data format.

How it was made

Admission check for every recording part: both streams present, stereo sync p99 ≤ 1 ms, dropped frames ≤ 10%, IMU continuity, duration ≥ 10 s.
Pre-screen at 1 Hz: exposure and sharpness gates, and at least one hand detected. Tasks where gloved hands near the frame edge defeat the detector were confirmed by hand.
Processing: stereo extrinsics refined from the scene when that lowers the rectified vertical disparity; 21-point hand keypoints detected in each view and triangulated.
Privacy: automatic face detection proposes draft boxes; every episode, both cameras, is then reviewed by a person who removes wrong boxes, adds missed faces and adjusts boxes over time, and only confirmed boxes are blurred. Audio removed; personal names and GPS removed; site names replaced by codes.

How each check is computed: Quality checks.

Known limitations

3D hand keypoints are model output, not motion capture. Accuracy against ground truth has not been measured.
No world-frame coordinates yet: the gravity-aligned frame fixes "up" from the IMU but moves with the head.
Left/right hand labels come from the 2D model and can be swapped.
Hands holding tools and gloved hands are detected less often.
Every episode was reviewed for faces by a person, but small or fast-moving faces can still be missed. Please report any missed face to contact@truetracedata.com.
Episodes are cut at quality boundaries, not task-cycle boundaries; there is no language annotation in this release.

License and commercial use

This sample is released under CC BY-NC 4.0 for non-commercial research. For commercial licensing or custom collection, contact contact@truetracedata.com.

Citation

@misc{truetrace_factory_stereo_ego,
  title  = {TrueTrace Factory Stereo Egocentric (Sample)},
  author = {TrueTrace Data Services},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/truetracedata/factory-stereo-egocentric}},
  note   = {Version v0.1}
}