Keypoint skeleton

Dots alone make it hard to tell which point belongs to which person. pf.annotate.keypoint_skeleton joins them with lines: shoulder to elbow, elbow to wrist, hip to knee.

Python
import pixelflow as pf
from ultralytics import YOLO

# The 17 COCO body keypoints, in the order the model returns them
COCO_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",
]
labels = [{"id": 0, "name": "person", "keypoints": [{"name": name} for name in COCO_KEYPOINTS]}]

model = YOLO("yolo11n-pose.pt")
detections = pf.from_ultralytics(model("street.jpg"), labels=labels)
detections = detections.filter_by_confidence(0.5)

image = pf.read_image("street.jpg")
image = pf.annotate.keypoint_skeleton(image, detections, min_confidence=0.5)
image = pf.annotate.keypoint(image, detections, min_confidence=0.5)
pf.save_image("street-skeletons.jpg", image)

Stick figures joining the shoulders, elbows, wrists, hips and knees of the people in a street photo

Keypoints need names

The skeleton connects points by name: ("left_shoulder", "left_elbow"). The model returns its points as a list without names, so tell the converter what they are, with labels in the form above: for each class, its id, its name and its keypoints in order. Without names, nothing is drawn and there is no error.

Your own connections

The default connections are the COCO body skeleton, for the 17 points above. For any other model, such as a hand or an animal, list the pairs to join, using the names you gave in labels:

Python
hand = [("wrist", "thumb_tip"), ("wrist", "index_tip"), ("wrist", "pinky_tip")]
image = pf.annotate.keypoint_skeleton(image, detections, connections=hand)

A pair is skipped when one of its points is missing, or when one of them scores below min_confidence.

Parameters

Parameter Default
image RGB image, drawn on in place
detections the Detections whose keypoints to join
connections the COCO body skeleton list of (name, name) pairs to join
thickness from the image size line width in pixels: 4 on a 1920×1280 image
colors the default palette list of RGB colors, picked by class_id (see colors)
min_confidence 0.0 skip a line unless both of its points score at least this