The COCO keypoint format: 17 keypoints, order, and skeleton.
Everything you need to read or write COCO person-keypoint annotations: the exact keypoint order, the official skeleton pairs, what the visibility flags mean, and worked JSON examples. Bookmark-grade; no scrolling through a tutorial to find the list.
What this format is, and where it comes from
COCO (Common Objects in Context) is the large-scale image dataset Microsoft released in 2014; its 2016 keypoint challenge labeled roughly 250,000 people with the same 17 body points, and the JSON layout of those annotations became the standard interchange format for 2D pose. Nearly every pose library reads or writes it, from pycocotools to OpenPose, Detectron2, MMPose, and YOLO. The layout is not person-specific (a category can define any names and skeleton), but the person schema below is what “COCO keypoints” means unqualified.
Keypoint order
The keypoints array is 0-based, but the skeleton pairs reference 1-based ids, the most common off-by-one in pose tooling. Both are listed. “Left” always means the subject’s left, which appears on the viewer’s right in a front-facing image.
| Array index (0-based) | Skeleton id (1-based) | Name | Side |
|---|---|---|---|
| 0 | 1 | nose | – |
| 1 | 2 | left_eye | left |
| 2 | 3 | right_eye | right |
| 3 | 4 | left_ear | left |
| 4 | 5 | right_ear | right |
| 5 | 6 | left_shoulder | left |
| 6 | 7 | right_shoulder | right |
| 7 | 8 | left_elbow | left |
| 8 | 9 | right_elbow | right |
| 9 | 10 | left_wrist | left |
| 10 | 11 | right_wrist | right |
| 11 | 12 | left_hip | left |
| 12 | 13 | right_hip | right |
| 13 | 14 | left_knee | left |
| 14 | 15 | right_knee | right |
| 15 | 16 | left_ankle | left |
| 16 | 17 | right_ankle | right |
Numbered with 1-based skeleton ids
The official 19 skeleton pairs
Each pair connects two keypoints by their 1-based ids. This is the canonical list from the COCO annotations, verbatim:
"skeleton": [ [16,14], [14,12], [17,15], [15,13], [12,13], [6,12], [7,13], [6,7], [6,8], [7,9], [8,10], [9,11], [2,3], [1,2], [1,3], [2,4], [3,5], [4,6], [5,7] ]
What v = 0, 1, 2 mean
Keypoints are stored as flat [x1, y1, v1, x2, y2, v2, …] triplets, 51 numbers for a person. The third value of each triplet is the visibility flag:
v = 0
Not labeled
x and y are 0; the point was not annotated at all.
v = 1
Labeled, not visible
The point has coordinates but is occluded (e.g. a hip under a coat).
v = 2
Labeled and visible
The point is annotated and visible in the image.
num_keypoints on the annotation counts the labeled points: those with v > 0. New to keypoint annotation itself? Start with the practical guide; looking for data in this format, see keypoint datasets.
The category and annotation JSON
The category declares the names and skeleton; each annotation carries the triplets. A minimal, valid pair:
// categories[]
{
"id": 1,
"name": "person",
"supercategory": "person",
"keypoints": [
"nose", "left_eye", "right_eye",
"left_ear", "right_ear",
"left_shoulder", "right_shoulder",
"left_elbow", "right_elbow",
"left_wrist", "right_wrist",
"left_hip", "right_hip",
"left_knee", "right_knee",
"left_ankle", "right_ankle"
],
"skeleton": [
[16,14],[14,12],[17,15],[15,13],
[12,13],[6,12],[7,13],[6,7],
[6,8],[7,9],[8,10],[9,11],
[2,3],[1,2],[1,3],[2,4],
[3,5],[4,6],[5,7]
]
}// annotations[]
{
"id": 42,
"image_id": 7,
"category_id": 1,
"num_keypoints": 3,
"keypoints": [
412, 143, 2, // nose: visible
431, 128, 2, // left_eye: visible
398, 129, 1, // right_eye: occluded
0, 0, 0, // left_ear: not labeled
// … one [x, y, v] triplet per
// keypoint, 17 in total
],
"bbox": [372, 94, 118, 340],
"iscrowd": 0,
"area": 40120
}Skip the conversion scripts.
DataTorch speaks this format natively: define the keypoints and skeleton on a label, annotate point-by-point in the browser with a live skeleton overlay, and import or export COCO keypoints, visibility flags included.
Tools and community for image annotation: label, review, and share datasets, from everyday photos to the imagery only specialists can read.
© 2026 DataTorch. All rights reserved.