Detect 33-point full-body pose landmarks in images and video with MediaPipe — normalized and world coordinates, visibility scores, and the POSE_CONNECTIONS skeleton helper.

Pose Estimation

The pose landmarker detects up to numPoses people in an image and returns 33 body landmarks per pose — covering the face, torso, arms, hands, legs, and feet. Each pose includes 3D world-coordinate landmarks and a confidence score.

Landmark Topology

MediaPipe's 33-point body topology spans the whole body:

Range	Region
0–10	Face (nose, eyes, ears, mouth)
11–16	Shoulders, elbows, wrists
17–22	Hands (pinky, index, thumb)
23–28	Hips, knees, ankles
29–32	Heels and foot index points

Each landmark is a { x, y, z, visibility } point. x and y are normalized to 0–1; z is depth; visibility (0–1) estimates the likelihood the point is visible and not occluded.

Model Variants

The catalog ships two pose models:

Catalog ID	Model	Size	Notes
`pose_landmarker`	Pose Landmarker (Lite)	5.8MB	Default — fast, good accuracy
`pose_landmarker_full`	Pose Landmarker (Full)	9.4MB	Higher accuracy, larger model

// Default lite model
const lite = mediapipe.poseLandmarker();

// Higher-accuracy full model
const full = mediapipe.poseLandmarker('pose_landmarker_full');

Detecting Poses

Create a model with mediapipe.poseLandmarker() and pass it to the core detectPose() function:

import { detectPose } from '@localmode/core';
import { mediapipe } from '@localmode/mediapipe';

const { poses, usage } = await detectPose({
  model: mediapipe.poseLandmarker(),
  image: imageBlob,
  numPoses: 1,
});

for (const pose of poses) {
  console.log(`Pose score ${pose.score.toFixed(2)} — ${pose.landmarks.length} landmarks`);
  const nose = pose.landmarks[0];
  console.log(`  nose at (${nose.x}, ${nose.y}), visibility ${nose.visibility}`);
}

console.log(`Detected in ${usage.durationMs.toFixed(0)}ms`);

Options

Option	Type	Default	Description
`model`	`PoseLandmarkModel`	—	The model from `mediapipe.poseLandmarker()`
`image`	`ImageInput`	—	Blob, `ImageData`, image/canvas/video element
`numPoses`	`number`	`1`	Maximum poses to detect
`minDetectionConfidence`	`number`	`0.5`	Minimum confidence threshold (0–1)
`abortSignal`	`AbortSignal`	—	Cancellation signal
`maxRetries`	`number`	`2`	Retry attempts on transient failure

Result

DetectPoseResult contains a poses array of PoseLandmarkResultItem:

interface PoseLandmarkResultItem {
  /** 33 pose landmarks in normalized image coordinates */
  landmarks: Landmark[];
  /** 33 pose landmarks in 3D world coordinates (meters, origin at body center) */
  worldLandmarks: Landmark[];
  /** Detection confidence score (0-1) */
  score: number;
}

Filter by visibility

Occluded joints still appear in landmarks but with a low visibility value. When drawing or measuring, skip points below a threshold (e.g. visibility > 0.5) for cleaner results.

Drawing the Skeleton

@localmode/core exports POSE_CONNECTIONS — the landmark index pairs that form the body skeleton:

import { detectPose, POSE_CONNECTIONS } from '@localmode/core';
import { mediapipe } from '@localmode/mediapipe';

const { poses } = await detectPose({
  model: mediapipe.poseLandmarker(),
  image: canvas,
});

const ctx = canvas.getContext('2d')!;

for (const pose of poses) {
  ctx.strokeStyle = '#00ddff';
  ctx.lineWidth = 4;
  for (const [start, end] of POSE_CONNECTIONS) {
    const a = pose.landmarks[start];
    const b = pose.landmarks[end];
    // Skip occluded joints
    if ((a.visibility ?? 1) < 0.5 || (b.visibility ?? 1) < 0.5) continue;
    ctx.beginPath();
    ctx.moveTo(a.x * canvas.width, a.y * canvas.height);
    ctx.lineTo(b.x * canvas.width, b.y * canvas.height);
    ctx.stroke();
  }

  ctx.fillStyle = '#ffffff';
  for (const point of pose.landmarks) {
    if ((point.visibility ?? 1) < 0.5) continue;
    ctx.beginPath();
    ctx.arc(point.x * canvas.width, point.y * canvas.height, 5, 0, Math.PI * 2);
    ctx.fill();
  }
}

React Hook

@localmode/react provides useDetectPose:

'use client';

import { useDetectPose } from '@localmode/react';
import { mediapipe } from '@localmode/mediapipe';

const model = mediapipe.poseLandmarker();

export function PoseDetector() {
  const { data, error, isLoading, execute, cancel } = useDetectPose({
    model,
    numPoses: 1,
  });

  return (
    <div>
      {isLoading && <button onClick={cancel}>Cancel</button>}
      {error && <p>{error.message}</p>}
      {data && <p>Detected {data.poses.length} pose(s)</p>}
    </div>
  );
}

The hook takes { model } (plus optional numPoses) and returns { data, error, isLoading, execute, cancel, reset }.

Real-Time Pose Tracking

For live webcam tracking, use createPoseTracker — it runs MediaPipe in VIDEO mode for higher throughput than calling detectPose() per frame. See the Streaming guide.

const tracker = mediapipe.createPoseTracker({
  video: videoElement,
  numPoses: 1,
  onResults: (poses, timestampMs) => drawPoses(poses),
});

await tracker.start();

Pose Estimation

Pose Estimation

Landmark Topology

Model Variants

Detecting Poses

Options

Result

Drawing the Skeleton

React Hook

Real-Time Pose Tracking

Next Steps

Hand Tracking

Real-Time Streaming

Face Detection

On this page