3D hand pose, per hand
Joint positions are annotated in 3D against the captured depth geometry, with a per-joint visibility flag, so occluded fingers are marked rather than guessed.
Video annotation
Egocentric video annotation.
First-person video only trains a policy once it carries the right labels. We annotate the footage we capture — hand pose, action segments, contact events, and instance masks — against the same synchronized depth and hardware clock the video was recorded on, then deliver it in the format your training stack reads.
- Labels
- 3D hand pose · action segments · contacts
- Basis
- Depth-aligned, not single-frame
- Formats
- RLDS · LeRobot · WebDataset
- License
- Buyer-owned
Definition
What is egocentric video annotation?
Egocentric video annotation is labeling first-person (head- or wrist-mounted) video with the structured signals a robot policy trains on — 3D hand pose, action-segment boundaries, contact and grasp events, and object or instance masks — tied to the synchronized depth stream and hardware clock the video was recorded on, rather than a caption or single classification tag per clip.
How it works
What egocentric video annotation actually covers.
Action segments, not clip-level tags
Each task is broken into start- and end-timestamped action segments rather than one label for the whole clip, so a policy can learn where one action ends and the next begins.
Contact and grasp events
Hand-object contact onsets and grasp events are annotated against the depth stream, giving a policy the moment of physical interaction, not just proximity in the frame.
Instance segmentation for the objects in scope
Object and body masks are annotated where the program spec calls for them, so a policy can separate the hand, the task object, and the background.
At a glance
What you get.
- Basis
- Synchronized RGB + depth + IMU
- Hand labels
- 3D joint pose · per-joint visibility
- Action labels
- Start/end-timestamped action segments
- Interaction labels
- Contact onsets · grasp events
- Optional
- Instance segmentation · de-identification
- Delivery
- RLDS · LeRobot · WebDataset · HDF5
Every collection runs through the same seven-stage end-to-end custom collection pipeline. End-to-end custom data collection is a managed service that takes an AI data need from problem to owned dataset in a single accountable pipeline.
FAQ
Egocentric video annotation, answered.
- What labels come with egocentric video annotation?
- 3D hand pose with per-joint visibility flags, start/end-timestamped action segments, and contact or grasp events as standard. Instance segmentation and de-identification (face and license-plate blurring) are added when the program spec calls for them.
- How is hand pose annotated in 3D rather than guessed from the image?
- Hand pose is annotated against the synchronized depth stream captured on the same hardware clock as the video, so joint positions reflect real geometry rather than being estimated from a single 2D frame.
- Can annotation be scoped to a specific skill or vocabulary?
- Yes. A skill spec sets the action vocabulary, environment, and label set before capture, so action-segment and contact labels use the terms your training stack expects rather than a generic taxonomy.
- What format is annotated data delivered in?
- RLDS, LeRobot, WebDataset, HDF5, or JSONL, matching the schema your training pipeline already reads, under a perpetual, buyer-owned license.
- Is annotation available on video we already have, or only on new capture?
- This program annotates video captured to your spec on our synchronized hardware, so hand pose and depth-based labels are grounded in the same clock the footage was recorded on.
Brief the gap. Get data back fast.
Tell us what your model is missing. We will quote reach, timeline, and price against a written spec, with a first validated batch in a median of 48 hours where coverage is deep.