Skip to content
annoroid
Free pilot

Process

From raw capture to a production dataset

Six steps, the same on every project. Each one leaves a record you can check, from the clean-up log to the quality report.
  1. 01Output · Data inventory

    Ingest

    Any capture, any rig.

    Robot rollouts, teleoperation, egocentric video, depth, LiDAR, IMU and joint states. We read your formats as they are (MCAP, ROS bags, HDF5, video plus logs) and inventory what's there.

  2. 02Output · Clean-up log

    Clean

    Only frames worth training on.

    Near-duplicates, blur, bad exposure, dropped frames and broken episodes are found and removed or repaired. Faces and plates are anonymized where needed.

  3. 03Output · Calibration report

    Correct geometry

    Right place, right scale, right time.

    Camera intrinsics and extrinsics are checked, depth and LiDAR are registered to the cameras, coordinate frames and units are unified, and every stream is aligned to one clock.

  4. 04Output · Labels

    Label

    Human-reviewed, frame by frame.

    Boxes, masks, tracks, 3D cuboids, hand and body keypoints, contacts and task captions. AI pre-labels where they help; a person reviews and corrects every item.

  5. 05Output · Quality report

    Validate

    Proof, not promises.

    Schema checks, cross-sensor consistency, track continuity and a second reviewer's spot-check on every batch, summarised in a quality report.

  6. 06Output · Production dataset

    Deliver

    Drops straight into training.

    COCO, YOLO, CVAT, LeRobot-style episodes, JSON or your own schema, with a dataset card. Fixes to anything we delivered are included.

Data processing

Raw capture is mostly repetition. We scored every frame of three raw 30-second clips for sharpness and duplication, removed what would only add noise, and normalised colour and exposure on what was left.

0→0frames in
worth labeling
  • Near-duplicates removed0
  • Blurred or transition frames removed0
  • Frames kept for labeling0
Raw frame and normalised frame — raw frame and labeled output
RAWLABELED

Kept frame · raw vs. colour and exposure normalised

48 consecutive frames, each marked kept, near-duplicate or blurred

48 consecutive frames (4.8 s) and their verdicts. Blur = least sharp 10% of each clip; duplicate = hash distance ≤ 2 from the last kept frame.

One raw clip in. A training-ready dataset out.

Inside a delivery

Estimated depth map of the pumpkin soup sceneDepth

Object tracks, hand-object contacts, timed captions and per-frame depth, packaged as a dataset your training code can load straight away.

What we deliver
Tomato salad · 21 tracks · 30 s

How it works

See the quality first. Then scale.

  1. Day 001

    Send a sample

    10 to 20 frames or a short clip, plus what you need labeled.

  2. ≤ 48h02

    Free pilot in 48 hours

    Labeled to your guidelines, with a quality report.

  3. Per item03

    Scale up

    Fixed per-image or per-minute pricing. No minimum contract.

  4. Every batch04

    Ongoing QA

    Every batch human-reviewed. Fixes included.

See your own data labeled before you pay anything.

Send 10 to 20 frames or a short clip. We label them to your guidelines and send a quality report within 48 hours.

Request a free pilot