Stereo Hands Sample: triangulated 3D hands on head-mounted stereo video
Twelve minutes of head-mounted stereo video of household and kitchen work, cut into 106 steps. Both hands are measured in 3D by stereo triangulation. The head is tracked from the same two views, so the hands are also given in a world frame that does not move with the head. It is in LeRobot v3 format and loads with the standard loader.
Preview: six steps, one from each source clip, with the triangulated hands drawn on both views.
The footage is not ours. It comes from the Humanola egocentric sample (CC BY 4.0). We ran our own open-source pipeline, ego2lerobot, on its two head cameras and added the labels described below. Humanola's own hand tracking and head pose are not copied here. We only used them to compare against.
Version 2 (16 November 2026). Added: head pose, hands in a world frame, the IMU. Changed: the calibration now has a lens distortion term that we measured from the footage (see the note below), so the 3D values moved by a few millimetres and some step boundaries moved. Version 1 had 109 steps.
What is in it
| Source | 6 of the 9 Humanola sample episodes: folding laundry, slicing carrots, peeling and cutting cucumbers, pouring cereal, washing dishes, peeling an apple |
| Video | Left and right head camera, 960×600 each, 15 fps (source: 1920×1200 at 30 fps) |
| Steps | 106 episodes, one per step, 2 to 15 seconds each, 12.0 minutes in total |
| Frames | 10,827 |
| Step names | 35 different names, written by a reviewer who looked at four frames of every step |
| Grasp and release events | 420, from the measured thumb to index distance |
| Head pose | Solved in more than 99.9% of frames, by stereo visual odometry; the world's y axis points along gravity |
For every frame:
observation.images.leftandobservation.images.right: the two views.observation.hands.kp2d.leftand.right: 21 keypoints per hand in each view, with a score.observation.hands.kp3d: 21 joints per hand in metres, in the left camera frame (x right, y down, z forward). Stored flat: left hand first, then right hand.observation.hands.valid: per hand, 1 if the hand is in 3D in this frame. When it is 0, that hand's 3D values, joint angles and wrist pose are stored as zeros.observation.hands.kp3d_status: per joint, 1 if triangulated in this frame, 2 if filled across a short gap, 0 if unknown.observation.hands.joint_angles: per hand, 15 finger bend angles, 4 spread angles and a closure value from 0 to 1. Angles are in radians.observation.hands.eeandaction: wrist position, wrist orientation and grip opening per hand. The action is the same vector one frame ahead.observation.head.pose: the left camera in the world frame. Position in metres, then the first two columns of the rotation (camera to world).observation.head.validis 1 where the pose was solved.observation.hands.kp3d_worldandobservation.hands.ee_world: the joints and the wrist pose in the world frame. They are zeros where the hand is not in 3D or the head pose was not solved.observation.imu: gyro (rad/s) and accelerometer (m/s²) of the head camera, averaged over the frame. The 200 Hz samples of each step are inmeta/imu/.observation.intrinsics.left,.rightandobservation.extrinsics: the calibration we used, at the stored image size.
In meta/: episodes_info.json (step name, source clip and frames, grasp events, hand coverage), calibration.json (one calibration per source clip), head_pose.json (the world frame and the tracking figures per clip), triangulation_stats.json and grasp_events.json. The quality report is at the top level (qc_report.html, .md, .json).
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("arete-robotics/stereo-hands-sample")
frame = ds[0]
frame["task"] # "fold a garment and put it on the stack"
frame["observation.hands.kp3d"].reshape(2, 21, 3) # left and right hand, 21 joints, (x, y, z) in metres
frame["observation.hands.joint_angles"].reshape(2, 20) # left and right hand, 20 values each
frame["observation.head.pose"] # x, y, z, then two columns of the rotation
frame["observation.hands.kp3d_world"].reshape(2, 21, 3) # the same joints in the world frame
How it was made
- MediaPipe finds both hands in the left and the right view.
- The two views are matched and each joint is triangulated. A hand is rejected when the two views disagree, or when its depth or size is not plausible.
- Short gaps are filled, joints are smoothed, and bone lengths are held near their median.
- Steps are cut where the motion pauses or a grasp or release happens.
- The head is tracked: scene points are triangulated in a keyframe and followed in the next frames, with the hands masked. The accelerometer is used only to point the world's y axis along gravity.
- A reviewer named every step from its frames.
No model wrote the step names. No depth network was used. Depth comes only from the two cameras.
A note on the calibration
The source card says the head pair is rectified and gives one set of intrinsics, no distortion and a 12 cm baseline. We measured two things in the files that do not fit that.
The rows do not line up. The right view sits about 45 pixels lower than the left view. With the published calibration alone, almost no hand passed the triangulation checks.
Some lens distortion is left. When we tracked the head, turns about the two image axes came out 9% to 18% smaller than the gyro's, while turns about the viewing axis were right. That pattern points to lens distortion. We measured one radial term from the footage alone, as the value where scene points triangulated in one frame land best in later frames. The six clips gave −0.020 to −0.038, and we use −0.030 for both lenses. With it the turn ratios rose to 0.92 to 0.97, and the rows of the two views agree better (90th percentile 1.6 to 2.4 pixels, from 2.2 to 4.6).
We kept the published focal length, principal point and baseline. We added the distortion term, and for the right view its vertical position, vertical scale and roll, fitted to background features matched between the views. The calibration for each clip is in meta/calibration.json.
This matters for anyone using the 3D values. The corrections are measured from the footage, not from a calibration target. The horizontal alignment of the right view cannot be measured this way, and an error there would shift all depths by a constant amount. The turn ratios below 1 say that the lens model is still not exact.
How good is it
Coverage. A hand is in 3D in 78% of frames for the left hand and 86% for the right hand. At least one hand is in 3D in 96% of frames. It varies by task: 95% or more when slicing carrots, and 58% to 64% when peeling an apple, where one hand often hides the other.
Jitter. The wrist leaves the line between its neighbouring frames by 2.6 to 2.7 mm at the median (1.7 mm in depth).
Comparison with another tracker. Humanola publishes its own 21 joint hand tracking for the same frames. It is not ground truth, but it is an independent system. Over 33,904 hand frames:
| Value | |
|---|---|
| Wrist distance between the two systems, median | 24.1 mm |
| Wrist distance, 90th percentile | 64.9 mm |
| Wrist within 3 cm / within 5 cm | 60% / 82% |
| All joints, median distance | 27.6 mm |
| Depth difference, median (ours is nearer) | 15 mm |
| Frame to frame wrist motion, ours / theirs | 3.6 mm / 3.1 mm |
Per clip and hand, the median wrist distance ranges from 12 mm to 46 mm. The wrists are about 0.55 m in front of the camera (median depth).
We cannot say which system is closer to the truth. Humanola's hands have the same palm length in every frame (9.5 cm), which suggests a fitted hand model. Ours measure 8.0 to 9.1 cm.
Head pose. The source ships a 200 Hz IMU, which gives two checks that do not depend on any camera tracking, and its own head pose, which gives a third. Figures are per clip, lowest to highest of the six.
| Ours | Head pose published with the footage | |
|---|---|---|
| Turn over one second against the integrated gyro, median | 0.26° to 0.73° | 0.36° to 1.91° |
| The same, 90th percentile | 0.5° to 1.6° | 1.2° to 4.1° |
| Tilt against the accelerometer over the whole clip, median | 0.4° to 2.4° | 0.4° to 2.2° |
| Tilt, 95th percentile | 1.8° to 6.1° | 1.5° to 6.9° |
The gyro check covers rotation over a short time. The tilt check covers drift in roll and pitch over a clip of 1.5 to 3 minutes. Nothing here checks position or heading against an independent reference. Against the published pose, a point 0.6 m in front of the camera moved differently by 2 to 18 mm over one second (median per clip). Our worst figure in every row is from the laundry clip, probably because large moving garments pull the tracking.
Limits
- Small. Twelve minutes. It is a sample to inspect, not a training set.
- Calibration is partly ours. See the note above.
- Occlusion. When one hand covers the other, the 3D hand is missing or filled.
- The head pose drifts. Nothing corrects it globally. It holds while the wearer stays at one work area and adds error when the view moves on. In the cereal clip tracking broke twice: poses there are comparable only within one part (
head_pose.partsinmeta/episodes_info.json). - The world frame is per source clip. Its origin is the left camera at the first frame of that clip.
- Step names are coarse. They were written from four frames per step by one reviewer.
- Three source episodes are not included (ironing, wiping, one of the two dish washing clips).
Bring your own video
Send us a few minutes of first-person video and we will label them for free, so you can judge the quality on your own footage. For a calibrated stereo pair you get triangulated 3D hands as in this sample. For a single camera you get 3D hands with depth estimated from hand size.
- What to send: a link to the original files, plus the calibration if it is stereo.
- What you get: a LeRobot v3 dataset, a preview video with the labels drawn on it, and a quality report.
- When: within 7 days.
- Your footage stays private. We use it only to label it for you.
Contact [email protected].
Licence and credit
This dataset is released under CC BY 4.0.
The video is from Humanola Egocentric Hand-Pose Dataset, Sample Delivery by Humanola Inc., CC BY 4.0: https://huggingface.co/datasets/humanola-inc/egocentric-sample
Changes made by Arete Robotics: the two head camera streams were downscaled to 960×600 and 15 fps and cut into steps. The hand keypoints, 3D joints, joint angles, wrist poses, head pose, world-frame hands, grasp events, step names, corrected calibration and quality report were added. The head camera's IMU is carried over. The wrist camera streams and Humanola's own tracking and head pose are not included.
The labels were made with ego2lerobot (Apache 2.0).
- Downloads last month
- 50