The 33 Body Landmarks Pose Estimation Tracks

Pose estimation is a model that looks at a video frame and predicts where your joints are. The common full-body set is 33 labelled points — face, shoulders, elbows, wrists, hands, hips, knees, ankles and feet — placed in every frame. Connect them and you get the skeleton overlay you see on swing analysis apps.

Key takeaways

• A "landmark" is a single named point with coordinates in the frame, plus a confidence value.
• The 33-point set breaks down as roughly 11 face points, 12 from the shoulders to the hands, and 10 from the hips to the feet.
• For golf, about ten of the 33 do nearly all the work: shoulders, elbows, wrists, hips, knees and ankles.
• One camera gives 2D positions; any depth is inferred, so rotations away from the lens are the least reliable measurements.
• Occlusion at the top of the backswing and loose clothing are real, unavoidable limits.

What is a body landmark in pose estimation?

A landmark is one named point on the body that the model locates in the image — "left wrist", "right hip" — returned as x and y coordinates within the frame, usually with a confidence score and sometimes an estimated depth. Run the model on every frame and each landmark becomes a track through time.

It is markerless. Nothing is attached to you; the model has learned from labelled images what a shoulder looks like and predicts where yours is. That is why lighting, contrast, clothing and camera angle affect the output so directly — they change the picture the model is reading.

The output is not a measurement in the way a tape measure is. It is a prediction, with an error attached, made independently for each frame. Good systems smooth those predictions across frames, which is one reason a higher frame rate helps: more samples, less noise per estimate. That argument is made in full in what 240 fps shows that 60 fps misses.

What are the 33 landmarks?

They cover the whole body, in three groups. Around eleven points sit on the face — nose, eye corners and centres, ears and mouth corners. Twelve run from the shoulders out through elbows and wrists to the hands. The remaining ten cover hips, knees, ankles, heels and toes.

Group	Points included	Roughly how many
Face	Nose, eyes and their inner/outer corners, ears, mouth corners	11
Arms and hands	Shoulders, elbows, wrists, plus thumb, index and pinky points on each hand	12
Lower body	Hips, knees, ankles, heels, foot index (toes)	10

The face group is generous because pose models are used for far more than sport. For golf it contributes one thing worth having: a stable reference for head position, so you can see head movement and whether you are staying behind the ball.

The hand points matter more than they look. Three points per hand mean the model can estimate the orientation of the hand, not just its location — which is the difference between knowing where your grip is and knowing what it is doing.

Which landmarks actually matter for a golf swing?

Ten of them carry nearly all the information: two shoulders, two elbows, two wrists, two hips and two knees, with the ankles close behind. Shoulders and hips give you turn and separation, elbows and wrists give you arm structure and release, and hips through ankles give you lower-body sequencing.

Shoulders and hips — turn, tilt and separation. The line between the shoulders and the line between the hips, tracked frame by frame, gives shoulder turn, hip turn and the difference between them. That difference is the X-factor, and it is the number most directly tied to how the body loads and unloads. See hip turn, X-factor and shoulder–hip separation.

Wrists and elbows — arm structure and release. The angle at the elbow tells you about width and about the trail arm folding and re-extending. The wrist positions, tracked against the elbow and the hand points, show when the hinge is held and when it is thrown. That is the mechanical basis of the casting versus lag distinction.

Hips, knees and ankles — sequencing. Sequence is about order: which segment starts down first, and how the motion passes upward. Ankles and knees anchor the frame of reference and reveal a slide, a sway or a stall long before any of it shows up in the arms.

Nose and ears — head stability. One reference point for head movement, which catches a reverse pivot or a lift out of posture.

The remaining points are not useless; they stabilise the model's estimate of the ones you care about. But if you are looking at a skeleton overlay wondering which lines to read, read those ten.

What can a single camera not see?

Depth. A 2D frame gives horizontal and vertical position directly and infers anything moving towards or away from the lens. Any measurement that depends on rotation in the third dimension — hip turn away from a face-on camera, for example — is an estimate the model derived rather than something it directly observed.

This is the fundamental limitation of markerless single-camera analysis, and no amount of frame rate or resolution removes it. It is also the practical reason to film both angles: down-the-line and face-on each measure directly what the other has to infer.

It is why camera placement carries so much weight. Put the lens on the target line and the club path is measured in the plane it happens in. Move it off the line and the model is now estimating a rotation it cannot see. The placement rules in how to film your golf swing exist to keep the interesting motion in the plane of the image.

Where does landmark tracking break down?

Three places, predictably: occlusion, clothing and light. At the top of the backswing the trail arm and both hands can pass behind the head and torso, so the model is predicting points it cannot see. Loose clothing moves independently of the limb inside it. Poor light removes the contrast the model needs at the edges.

• Occlusion. The top of the backswing and the finish are the worst frames in any swing. Models fill in occluded points from context and the surrounding frames — a reasonable estimate, but an estimate. Confidence values usually drop there, and any system worth using tells you so rather than hiding it.
• Clothing. A baggy jacket makes a shoulder point wander. A fitted top does not. This is the cheapest accuracy improvement available to you and it costs nothing.
• Light and background. Backlighting turns you into a silhouette and the wrists go first. A cluttered background gives the model limb-shaped distractions.
• The club. The 33 landmarks are body points. The club is not among them — tracking a shaft and a clubhead is a separate problem, solved separately, and worth knowing about when an app shows you a club path line.

We are deliberately not quoting accuracy figures here. Published numbers depend on the specific model, the dataset and the conditions, and none of them would tell you what your phone will do at your range in your light. What is honest to say is the direction of the effects: contrast helps, occlusion hurts, and stable camera geometry matters more than any of it.

How does this turn into a swing measurement?

By geometry. Once every frame has labelled points, angles become subtraction. Shoulder turn is the angle of the shoulder line against address. Elbow angle comes from three points. Head movement is the displacement of one point over time. The skeleton is the raw data; the numbers are what you compute from it.

That is the whole pipeline in one sentence: frames in, landmarks per frame, angles from landmarks, and then a comparison of those angles against the checkpoints of the swing. Those checkpoints have names — P1 through P10 — and they are where a skeleton overlay stops being a novelty and starts being a lesson.

SwingOK is built on exactly this: 33 landmarks on every frame of a high-frame-rate clip, turned into the handful of angles that actually change how you strike the ball.

Where should you go next?

The skeleton is only as good as the video it was extracted from, so the camera setup comes first. From there, the pieces below cover the frame-rate arithmetic, the specific angles the landmarks make measurable, and the positions worth checking in each frame.

• How to film your golf swing — the pillar: angle, placement, lens height, light and background.
• What 240 fps shows that 60 fps misses — why the number of frames changes what the landmarks can tell you.
• Hip turn, X-factor and shoulder–hip separation — the measurement that shoulder and hip landmarks exist to produce.
• The golf swing sequence: P1–P10 explained — the checkpoints the angles get compared against.