AI Pose Detector

What it draws

Upload a photo of a person and the page overlays a red skeleton: a point at each tracked joint, connected into limbs along the body, the arms, the legs and the face. The panel reports 33 landmarks detected, and the annotated photo stays on the canvas.

Unlike most of the other AI tools in this group, this one has no download button. The result is a visual overlay you inspect on screen; to keep it, you would need a screenshot. There is also no measurement tool, no angle readout and no export of coordinates.

The model

The "lite" variant is the fastest of the Pose Landmarker family. It is accurate enough for form checks, silhouette work and rough animation reference, but it is less precise around hands, feet and heavily foreshortened limbs than the larger variants. Choosing it is what makes the page usable on an ordinary laptop without a GPU.

The single-pose configuration means one person per image. In a photo with two people, the model locks onto one — usually the larger or more central figure — and ignores the other.

The 33 points

The landmark set covers the body in a fixed order: nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles, heels and toe tips, plus a handful of facial points. That is enough to describe the overall posture and the main limb angles, and not enough to describe fingers, spine curvature or the sole of the foot in detail.

Points are drawn as white dots with a red outline, joined by red lines where the model defines a connection. Because the canvas is sized to the image's natural dimensions, the skeleton lines up with the body at full resolution. The interface does not show landmark names or indices, so the overlay is descriptive rather than analytical.

Step by step

1. Click the upload area and choose a photo. Drag and drop is not supported.
2. Wait while the runtime and the Lite model download on first use. The panel names each stage.
3. Inference runs automatically after the image loads; there is no separate button.
4. Check the skeleton. A good fit has joints centred on the body and limbs following the visible pose.
5. If you need to keep the result, take a screenshot — the page does not offer a download.

The Load AI Model button warms up the model ahead of time and changes to "Model Ready" when it finishes. Replacing the image re-runs detection with the already loaded model, which is near-instant.

Where it is genuinely useful

• Fitness form review. A skeleton shows whether knees track over toes, whether hips are level, and how deep a squat actually is.
• Dance and sports reference. Extract a pose from a photo to plan a drawing or a keyframe.
• Ergonomics. See the relationship between head, shoulders and hips in a desk setup photo.
• Yoga and physiotherapy self-checks. A quick visual confirmation that a posture matches the intended shape.
• Prototyping. Check whether a camera angle captures enough of the body before building a live-tracking app.

It is a visual aid, not a coach. It does not know the exercise, does not count repetitions, does not judge safety and cannot tell a good lift from a dangerous one. Anyone using it for training should treat it as a mirror with dots, not as an instructor.

Where it fails

Occlusion is the main problem. A person behind a desk, a limb out of frame, or a body crossed by another object leaves the model inferring positions that may be visibly wrong. Loose clothing hides joint locations, and so do heavy shadows.

Extreme foreshortening — a limb pointing at the camera — compresses a whole segment into a few pixels, and the skeleton can fold or shorten unnaturally. Unusual poses, such as a handstand or a deep twist, are outside the common training distribution and often produce tangled connections.

Small figures are another limitation. A person occupying a small part of a wide landscape gives too few pixels for reliable joints. Very low resolution, motion blur and watermarks across the body all reduce quality.

Because the output is 2D, a pose that looks correct from the camera's viewpoint may be anatomically impossible from another angle. The model does not tell you that; it simply places points in the image plane.

Preparing a photo

The best input is a full-body or three-quarter view with the whole person inside the frame, taken from roughly waist height rather than from above or below. Keep the limbs clear of furniture and other people, avoid baggy clothing that hides the hips and knees, and make sure the light is even enough to see the silhouette. If the subject is small in the frame, crop first so the body fills more of the image; the model has more pixels to work with and the joints land more accurately.

What it does not do

No multi-person tracking, no video, no webcam, no repetition counting, no angle computation, no 3D model export, no coordinate download and no annotated image download. There is no comparison between two poses and no measurement in centimetres.

It also does not identify anyone. Pose landmarks are not biometric identifiers in the way a face is, but they are still personal data when tied to an individual, and the same consent considerations apply to photos of other people.

Privacy

The image is processed locally in WebAssembly. It is decoded in the browser, passed to the model, and drawn to a canvas; nothing is uploaded and nothing is stored. The page downloads its runtime and weights from public CDNs on first use, which reveals that the tool was loaded but not what was in the photo.

Limits

One image, one person, one inference pass, no download, no batch, no video and no webcam. No threshold or confidence control, no landmark table, no measurement tools and no multi-pose support. Large images take longer to decode, and the page needs a modern browser with WebAssembly plus network access on first load.

Common mistakes

• Uploading a group photo. Only one pose is tracked.
• Looking for a download button. There is none; the overlay is on-screen only.
• Expecting joint angles. The page draws points and lines, not measurements.
• Using a distant subject. Small figures give poor landmarks.
• Trusting a foreshortened limb. Depth is lost in a 2D projection.
• Treating it as a form coach. It has no idea what exercise is being performed.

Frequently Asked Questions