The Pose contract

A Pose is one frame of body points. The whole engine trusts that poses follow this contract; if you write your own PoseProvider, it’s up to you to respect it.

public struct Pose: Sendable {
    public var points: [Joint: CGPoint]
    public var timestamp: TimeInterval
    public var imageSize: CGSize?
    public var confidence: [Joint: Double]
}

The four rules

1. Image pixels, with y pointing down. The origin is at the top left, as in UIKit. Don’t use normalized coordinates (0…1) or Vision’s bottom-left origin.
2. Anatomical sides. .leftElbow is the patient’s real left elbow. With the front camera the image is mirrored, so the adapter must swap left and right before building the pose.
3. Only confident points. Low-confidence joints are left out of the dictionary. A missing joint means “I can’t see it”, and the engine treats it that way: it doesn’t measure, doesn’t penalize and reports it as missing.
4. confidence is optional. If your detector has no confidence, leave the dictionary empty; confidence metrics will be nil.

Empty poses

Send the pose even when it has no points. An empty pose tells the session the patient left the frame, so it can emit .visibilityChanged.

The 19 joints

Joint follows the 19 points of Vision’s 2D body detector:

Area	Joints
Head	nose, leftEye, rightEye, leftEar, rightEar
Midline	neck, root (center of the hips)
Arms	leftShoulder, rightShoulder, leftElbow, rightElbow, leftWrist, rightWrist
Legs	leftHip, rightHip, leftKnee, rightKnee, leftAnkle, rightAnkle

To avoid writing side == .left ? .leftKnee : .rightKnee, use the shortcuts:

Joint.knee(.left)       // .leftKnee
Joint.shoulder(side)    // depends on the side
BodySide.right.opposite // .left

Note

There are no points for toes or fingers. That’s why ankle dorsiflexion is measured as shin tilt; see the catalog.

Reading a pose

if let elbow = pose[.rightElbow] {
    print(elbow)                                // CGPoint in pixels
}
pose.confidence(of: [.rightShoulder, .rightElbow, .rightWrist])
// The lowest confidence among the visible ones, or nil

imageSize lets you draw the points over the camera view by scaling them to the view’s size.