Results & types

PoseDetectionResultBundle

What onResults and PoseDetectionOnImage give you.

interface PoseDetectionResultBundle {
  results: PoseLandmarkerResult[];
  inferenceTime: number; // milliseconds
  inputImageWidth: number;
  inputImageHeight: number;
}

interface PoseLandmarkerResult {
  landmarks: Landmark[][];      // [person][landmarkIndex], normalized
  worldLandmarks: Landmark[][]; // [person][landmarkIndex], meters
  segmentationMasks: Mask[];
}

For live camera and single images, results has one entry. So the usual way in is:

const people = result.results[0]?.landmarks ?? [];
const firstPerson = people[0]; // 33 landmarks, or undefined if nobody is in frame

Landmark

interface Landmark {
  x: number;
  y: number;
  z: number;
  visibility?: number; // 0–1, how likely the point is visible (not hidden)
  presence?: number;   // 0–1, how likely the point is in the frame at all
}

• In landmarks, x and y are normalized to the input image (0 to 1). z is depth relative to the hips; smaller is closer to the camera.
• In worldLandmarks, x, y and z are real-world 3D coordinates in meters, with the origin between the hips. Use these for angles and distances that shouldn’t change with how far someone stands from the camera.

ViewCoordinator

Passed as the second argument to onResults.

interface ViewCoordinator {
  getFrameDims(info: { inputImageWidth: number; inputImageHeight: number }): { width: number; height: number };
  convertPoint(frame: { width: number; height: number }, p: { x: number; y: number }): { x: number; y: number };
}

See Drawing a skeleton overlay.

DetectionError

interface DetectionError {
  code: number;
  message: string;
}

Enums

enum Delegate {
  CPU = 0,
  GPU = 1,
}

enum RunningMode {
  IMAGE = 0,
  VIDEO = 1,
  LIVE_STREAM = 2,
}

Other exports

Export	Description
KnownPoseLandmarks	Landmark name → index map. See Landmarks.
KnownPoseLandmarkConnections	Pairs of landmark indices that form the skeleton.
BaseViewCoordinator	The class behind ViewCoordinator, if you need to build one yourself.
MediaPipeSolution, DetectionCallbacks, Mask, ResizeMode, ImageOrientation	Supporting types.