MediaPipe pose landmarks: all 33 by index

The pattern is easy to keep in your head past the face: from 11 onwards, odd indices are the subject's left and even indices their right, working down the body — shoulders 11/12, elbows 13/14, wrists 15/16, hands 17–22, hips 23/24, knees 25/26, ankles 27/28, heels 29/30, toes 31/32.

Normalized coordinates are convenient and also the most common source of wrong angles. On a portrait 1080×1920 frame, one unit of x is 1,080 pixels and one unit of y is 1,920, so a segment's direction is distorted until you multiply back by the frame size. How to calculate joint angles covers why that matters.

The pose landmarker returns two sets for each person. Image landmarks (pose_landmarks in Python, landmarks in JavaScript) are the normalized positions described above. World landmarks (pose_world_landmarks, worldLandmarks) are the same 33 points as 3D coordinates in metres, with the origin between the hips.

• Use image landmarks when you are drawing on the video, or measuring an angle in the plane the camera sees — which is what a side-on or front-on movement analysis wants.
• World landmarks are tempting for angles out of the camera's plane, but their depth is inferred from one 2D image. Check them against what you can see before you rely on them.

left_knee is the knee on the person's left, not the one on the left of the picture. Facing the camera, the subject's left side appears on the right of the image. That is the right way round — until a front camera mirrors the preview. Mirror the image, and the model still labels the anatomy correctly, but the left knee is now drawn where the viewer expects the right one. Decide once whether you mirror, and do it after pose estimation, not before.

Every mainstream pose model uses the subject's own left and right: MediaPipe, ML Kit, MoveNet and Apple Vision alike. The engines compared has the full mapping between them.

import mediapipe as mp
from mediapipe.tasks import python
from mediapipe.tasks.python import vision

options = vision.PoseLandmarkerOptions(
    base_options=python.BaseOptions(model_asset_path="pose_landmarker_full.task"),
    running_mode=vision.RunningMode.IMAGE,
)

with vision.PoseLandmarker.create_from_options(options) as landmarker:
    image = mp.Image.create_from_file("frame.jpg")
    result = landmarker.detect(image)

    if result.pose_landmarks:                # one list per detected person
        person = result.pose_landmarks[0]    # 33 landmarks, index 0–32
        knee = person[25]                    # left_knee
        print(knee.x, knee.y, knee.z, knee.visibility)

import { FilesetResolver, PoseLandmarker } from "@mediapipe/tasks-vision";

const fileset = await FilesetResolver.forVisionTasks(
  "https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision/wasm"
);
const landmarker = await PoseLandmarker.createFromOptions(fileset, {
  baseOptions: { modelAssetPath: "pose_landmarker_lite.task" },
  runningMode: "VIDEO",
});

// Once per video frame:
const result = landmarker.detectForVideo(video, performance.now());
const person = result.landmarks[0];      // undefined when nobody is in frame
const leftKnee = person?.[25];           // { x, y, z, visibility }

The older Solutions API (mp.solutions.pose) returns the same 33 points in the same order, and names them in the PoseLandmark enum — PoseLandmark.LEFT_KNEE is 25 — which is worth using in place of bare numbers once the code grows.

The twelve points from shoulders to ankles are common to all of them, which is why a rep counter built on those twelve runs on any engine. Hands, heels and toes are MediaPipe's own: a wrist or ankle angle needs them.

Upload a clip to Markerless and every frame's landmarks come back with it: drawn over the video so you can check what the model saw, and downloadable as JSON or CSV with the frame size recorded, so the pixel conversion is one multiplication away. The same landmarks give fourteen joint angles per frame, and a rep counter you can export to whichever engine your app runs.