Skip to content
Annoroid
Request a free pilot
All work

Sample project

Overhead kitchen task: 20 object tracks, hand contacts and action captions

TrackingActions & CaptionsHand-object
Overhead kitchen task: 20 object tracks, hand contacts and action captions — raw frame and labeled output
RAWLABELED
Data
Video · 1920×1080
Volume
150 frames · 2,711 labeled instances
Status
Sample project

The problem

A robot learning a kitchen task needs to know what every object is, where it goes from frame to frame, when a hand touches it, and which step of the task is happening. Raw video has none of that.

What we did

  • Object inventory of 19 classes, from knife and peeler to recipe card
  • Anchor boxes placed by a reviewer, then propagated with SAM 2.1 video segmentation
  • 20 tracks reviewed on per-object contact sheets; drifting tracks corrected with extra anchors
  • 23 hand-object contact events derived from mask overlap and reviewed
  • 10 timestamped action captions plus scene captions
  • Exported as COCO (with outlines), YOLO, CVAT video XML and WebVTT

Result

150 frames with 2,711 labeled instances across 20 tracks, 23 hand-object contacts and 12 captions, delivered in four formats.

Labeled video · 30 seconds

  1. Caption track
  2. 00:00Yellow watermelon wedges are cut into cubes on the cutting board.
  3. 00:04A glass mixing bowl is set on the board and the watermelon cubes are moved into it.
  4. 00:08The knife is set down and the cucumber is brought to the board.
  5. 00:09The cucumber is peeled with the red peeler.
  6. 00:11The peeled cucumber is halved lengthwise and diced with the knife.
  7. 00:15The diced cucumber is added to the mixing bowl and the last pieces are cleared from the board.
  8. 00:16The board is wiped with a paper towel.
  9. 00:17The red onion is picked up, peeled and cut in half.
  10. 00:20The onion half is sliced and finely chopped with the knife.
  11. 00:29Chopped onion is scraped off the board into the mixing bowl with the knife.

Frames, depth and review sheet

Labeled frame 18
Frame 18 · labeled
Labeled frame 90
Frame 90 · labeled
Labeled frame 126
Frame 126 · labeled
Estimated depth for the same scene
Depth · frame 52
Contact sheet of 15 labeled frames used to review every track
Review contact sheet · every 10th frame

Sample output

An excerpt from the delivered files. "source": "ai" marks boxes propagated by the segmentation model; every one was reviewed by a person before export.

kitchen.jsonexcerpt
{
  "image_id": 55,
  "category": "olive oil bottle",
  "bbox": [63, 10, 273, 209],
  "attributes": {
    "track_id": "obj_5fe3a3683a8a",
    "occluded": false,
    "keyframe": true,
    "source": "ai"
  },
  "segmentation": [[69, 29, 64, 40, 65, 58, 90, 88, "…"]]
}

Source footage: “Farmers Market Food Demo - Watermelon Salad”, US Department of Agriculture (Agricultural Marketing Service), via Wikimedia Commons (Public domain). Labels, captions, tracks and depth by Annoroid.

Request a free pilot for your own data.

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