MediaPipe Pose: The Complete 33 Landmark List (Index Reference)

Published: May 2, 2026
For: Developers using MediaPipe / anyone implementing pose estimation

This article is for you if:

  • You want the index and location of every MediaPipe Pose landmark
  • You only need a few specific joints but don't know their indices
  • You want to know how to use the visibility score

TL;DR

  • MediaPipe Pose returns 33 landmarks (11 on the face, 12 upper body, 10 lower body)
  • Each landmark has four values: { x, y, z, visibility }
  • Joints with visibility below 0.5 are low-confidence and best skipped
  • Left/right are from the person's point of view, not the camera's (as if mirrored)

Introduction

When you implement MediaPipe Pose, you keep asking yourself "which index is the shoulder again?" or "is 15 the left or the right?" The official docs have a diagram, but opening it mid-implementation just to double-check is a small hassle.

This article lays out all 33 landmarks in one place, explains how to pick the joints you use most, and shows how to use visibility. It also includes a real example from Pose Mirror.

Landmark list

Full-body layout (text diagram)

          0 (nose)
    1,3 ---- 2,4  (eyes: inner / outer)
    5          6  (eye outer corners)
    7          8  (ears)
    9         10  (mouth: left / right)

        11 --- 12  (shoulders)
       /         \
     13           14  (elbows)
      |             |
     15           16  (wrists)
   17,19,21   18,20,22  (fingers)

        23 --- 24  (hips)
       /         \
     25           26  (knees)
      |             |
     27           28  (ankles)
   29,31       30,32  (heels / toes)

Reading left vs. right: odd indices (11, 13, 15...) are the person's left side; even indices (12, 14, 16...) are their right side. Viewed from the front, this is mirror-reversed relative to the camera's view.

Face / head (0–10)

Index Part Notes
0 Nose Face center; useful for orientation
1 Left eye (inner)
2 Right eye (inner)
3 Left eye (outer)
4 Right eye (outer)
5 Left eye (outer corner)
6 Right eye (outer corner)
7 Left ear Often low-accuracy in profile
8 Right ear
9 Mouth (left)
10 Mouth (right)

Upper body (11–22)

Index Part Common pairing
11 Left shoulder 11→13 (left upper-arm direction)
12 Right shoulder 12→14 (right upper-arm direction)
13 Left elbow 13→15 (left forearm direction)
14 Right elbow 14→16 (right forearm direction)
15 Left wrist
16 Right wrist
17 Left hand (index knuckle) When you need hand detail
18 Right hand (index knuckle)
19 Left hand (pinky knuckle)
20 Right hand (pinky knuckle)
21 Left thumb
22 Right thumb

Lower body (23–32)

Index Part Common pairing
23 Left hip 23→25 (left thigh direction)
24 Right hip 24→26 (right thigh direction)
25 Left knee 25→27 (left shin direction)
26 Right knee 26→28 (right shin direction)
27 Left ankle
28 Right ankle
29 Left heel
30 Right heel
31 Left toe
32 Right toe

Accessing landmarks in code

pose.onResults((results) => {
  const lm = results.poseLandmarks;
  if (!lm) return; // no person detected

  // Examples for commonly used joints
  const leftShoulder  = lm[11]; // { x, y, z, visibility }
  const rightShoulder = lm[12];
  const leftElbow     = lm[13];
  const nose          = lm[0];

  // Check confidence with visibility before using it
  if (leftShoulder.visibility > 0.5) {
    console.log(
      `Left shoulder: x=${leftShoulder.x.toFixed(3)}, y=${leftShoulder.y.toFixed(3)}`
    );
  }
});

What the coordinates mean:

  • x, y: normalized to the image size (0–1). Top-left is (0,0), bottom-right is (1,1)
  • z: depth relative to the hips (negative is toward the camera). Less accurate than x, y
  • visibility: detection confidence for the joint (0–1)

Using visibility correctly

Each landmark's visibility indicates how reliably that joint is captured.

// BAD: ignore visibility and use it as-is (mixes in low-accuracy joints)
const shoulder = { x: lm[11].x, y: lm[11].y };

// GOOD: only use high-confidence joints
if (lm[11].visibility > 0.5) {
  const shoulder = { x: lm[11].x, y: lm[11].y };
  applyToModel(shoulder);
}

Cases where visibility tends to drop:

Situation Affected joints
Side-on or back-facing photos All joints on the hidden side
Limbs overlapping the body The overlapping limbs
Part of the body cropped off-screen Joints in the cropped area
Dark or blurry images Generally lower across the board

A real example from Pose Mirror

Pose Mirror doesn't use all 33 points — it narrows them to the 12 points that map to the 3D figure's joints.

// The joints Pose Mirror actually uses (12 points)
const jointMap = {
  leftShoulder:  lm[11],
  rightShoulder: lm[12],
  leftElbow:     lm[13],
  rightElbow:    lm[14],
  leftWrist:     lm[15],
  rightWrist:    lm[16],
  leftHip:       lm[23],
  rightHip:      lm[24],
  leftKnee:      lm[25],
  rightKnee:     lm[26],
  leftAnkle:     lm[27],
  rightAnkle:    lm[28],
};

The face (0–10) isn't used for head orientation, and fine finger motion (17–22) isn't supported yet. Bringing those in would enable more precise pose transfer, including head direction and fingertips.

Summary

  • Landmarks use fixed indices 0–32
  • Left/right are from the person's viewpoint (as if mirrored)
  • Skip joints with visibility < 0.5 — they're low-confidence
  • The key joints for 3D pose transfer are the 18 points from 11 to 28 (limb roots to ends)

Try Pose Mirror and see how the 33 landmarks get turned into a 3D pose for yourself.

Related (Japanese):

References

← Back to portfolio TOP