| name | mediapipe-usage |
| description | Summarizes Google MediaPipe usage for web: Pose Landmarker with @mediapipe/tasks-vision, landmark indices, running modes, and patterns for real-time video. Use when working with MediaPipe, pose detection, body landmarks, or @mediapipe/tasks-vision. |
Google MediaPipe Usage (Web / Pose Landmarker)
When to Use This Skill
- Using or debugging
@mediapipe/tasks-vision, Pose Landmarker, or pose detection in a web app.
- Needing the 33 pose landmark indices, skeleton connections, or mapping landmarks to gestures or control signals.
Setup
Install (prefer pnpm):
pnpm add @mediapipe/tasks-vision
WASM root: Resolve vision tasks from CDN when creating the task:
const vision = await FilesetResolver.forVisionTasks(
"https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@latest/wasm",
);
Create the Pose Landmarker Task
Use PoseLandmarker.createFromOptions(vision, options):
import { PoseLandmarker, FilesetResolver } from "@mediapipe/tasks-vision";
const vision = await FilesetResolver.forVisionTasks(
"https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@latest/wasm",
);
const poseLandmarker = await PoseLandmarker.createFromOptions(vision, {
baseOptions: {
modelAssetPath: modelUrl,
delegate: "GPU",
},
runningMode: "VIDEO",
numPoses: 1,
minPoseDetectionConfidence: 0.5,
minPosePresenceConfidence: 0.5,
minTrackingConfidence: 0.5,
});
- runningMode:
IMAGE for single image → use detect(image). VIDEO for stream → use detectForVideo(video, timestamp).
- baseOptions.modelAssetPath: URL to a
.task model (lite / full / heavy). See reference.md for URLs.
- delegate:
"GPU" preferred; some environments fall back to CPU.
Run the Task
Single image (runningMode IMAGE):
const result = poseLandmarker.detect(imageElement);
Video / webcam (runningMode VIDEO):
Call detectForVideo(video, timestamp) inside a requestAnimationFrame loop. Throttle by time (e.g. ~33 ms between frames) to avoid excessive work:
let lastFrameTime = 0;
function detectLoop() {
const now = performance.now();
if (video.readyState >= 2 && now - lastFrameTime > 33) {
lastFrameTime = now;
const result = poseLandmarker.detectForVideo(video, now);
if (result.landmarks?.length) {
const landmarks = result.landmarks[0];
}
}
requestAnimationFrame(detectLoop);
}
requestAnimationFrame(detectLoop);
Result Shape
- result.landmarks: Array of poses; each pose is
NormalizedLandmark[] (33 points). Each landmark: x, y, z (normalized 0–1; z is depth relative to hip center), visibility (0–1).
- result.worldLandmarks: Optional 3D coordinates in meters (same indices).
- Single person: use
result.landmarks[0].
Practical Patterns (Know-how)
- State machine: idle → loading (load model) → ready (can start) → active (webcam + detection) → error. When switching model variant, close the old PoseLandmarker instance and create a new one.
- Throttle: Run
detectForVideo only when performance.now() - lastFrameTime > 33 (≈30 fps) to avoid blocking the main thread.
- Smoothing: Apply a smoothing factor (e.g. 0.3) to derived values (pitch, bank) to reduce jitter; use a dead zone (in degrees) to ignore small movements.
- Confidence: Use each landmark’s
visibility; ignore or downweight points below a threshold. Helper: getLandmark(landmarks, index, minConfidence) returning the point only if visibility >= minConfidence.
- Gestures: e.g. “hands forward” = compare shoulder vs wrist z; “hands overhead” = compare wrist y to shoulder y. Use consecutive-frame counters for toggles (e.g. require N frames in pose before firing an action).
Cleanup
- Stop webcam:
stream.getTracks().forEach(t => t.stop()).
- Release task:
poseLandmarker.close() when done or before creating a new instance.
Additional Resources