AI Holistic Tracker

Three trackers in one pass

Upload a photo of one person and the page marks the face, the body pose and both hands at the same time. It reports the total number of landmarks found — up to 553 when everything is visible — and draws them as coloured dots over the image.

This is a combined view rather than a new model. MediaPipe's Holistic Landmarker runs a face mesh, a pose estimator and a two-hand tracker together, then returns all three results from a single image. The page draws each set in a different colour and adds up the points.

What is being counted

The face dominates the count, which is why the panel usually shows a number in the high five hundreds for a clear portrait. A missing hand costs 21 points, a missing pose costs 33, and a face that is turned away can cost most of the 478.

The model is configured with face blendshapes enabled, but the interface does not display or export them. There is no expression readout, no head-pose value and no gesture label.

What the overlay shows — and what it does not

Every landmark is drawn as an independent dot. There are no connecting lines between them, so the body does not appear as a skeleton and the hands do not appear as finger diagrams. The face becomes a dense green stipple, the pose a sparse outline of joints, and the hands two small clusters.

That design is useful for seeing raw coverage: you can tell at a glance which parts the model found. It is less useful for reading a pose, because the connections that make a skeleton legible are absent. The overall silhouette is still implied by the dots, but you have to reconstruct it yourself.

The canvas is set to the photo's natural dimensions, so the dots land on the correct pixels at full resolution. There is no download button; the result is for on-screen inspection, and keeping it means taking a screenshot.

Step by step

1. Click the upload area and choose a photo with one person in it. Drag and drop is not supported.
2. On first use, the browser downloads the runtime and the holistic model. The panel names each stage.
3. Detection runs automatically after the image loads, with no button to press.
4. Read the total and check which colour groups are present. Missing colours tell you which part was not found.
5. Upload another photo to run again; the loaded model makes repeat runs fast.

The Load AI Model button preloads everything ahead of time and changes to "Model Ready" when it succeeds.

Where it is useful

• Animation reference. Capture face, body and hand positions from one photo to plan a character pose.
• Movement analysis. See the relationship between head, shoulders, hips and hands in a sports or dance frame.
• AR prototyping. Confirm that a camera angle captures all three regions before building a live filter.
• Ergonomics and posture. Check head position relative to the shoulders and hands.
• Teaching. A clear demonstration of how several landmark models can be combined in one pipeline.

It is not a form coach, a physiotherapy tool or a motion-capture system. It gives positions in a single 2D image, with no timing, no depth and no measurement.

Where it fails

Occlusion is the main limitation. A hand behind the body, a face turned away or an arm out of frame removes the evidence the model needs, and the missing points simply do not appear. Unlike some pages, there is no warning per region — only a lower total.

Small figures are another weakness. A full-body shot taken from far away leaves too few pixels for the face mesh and the hands, so the total collapses toward the pose count. Cropping closer usually restores the missing groups.

Loose clothing hides the hips and knees; heavy shadow hides the face; motion blur softens every edge. Unusual poses — a handstand, a deep twist, a limb pointing at the camera — produce joints that look anatomically odd because the model is projecting a 3D structure onto a 2D image.

Two people in the frame is a problem too. The holistic model is built for one subject, so it tracks one person and may switch between them if both are prominent. The page has no way to choose which one.

Reading the total

553 is the ceiling, not a quality score. A clean, front-facing, full-body photo with both hands visible will approach it. A portrait with hands out of frame will sit around 511. A distant figure may show only the 33 pose points. Use the number to spot missing regions, not to compare the accuracy of two different photos.

Because the dots are drawn without confidence information, a guessed landmark looks the same as a confident one. Check the overlay against the visible anatomy: points that drift onto the background or bunch at an occluded edge are guesses.

What it does not do

No connections or skeleton lines, no per-region counts, no blendshape display, no gesture names, no download, no coordinate export, no video, no webcam, no multi-person tracking and no 3D export. It does not measure joint angles, count repetitions or identify anyone.

Privacy

The image is decoded and processed locally in WebAssembly. Nothing is uploaded and nothing is stored after the page is closed. The runtime and model weights are fetched from public CDNs on first load, which reveals that the tool was opened but not what was in the photo.

Face and hand geometry are sensitive when tied to a person. Local processing solves the transmission question, not the consent question, so handle photos of other people accordingly.

Limits

One image, one person, one inference pass, no download and no numeric output. No threshold control, no way to toggle individual regions and no batch mode. Very large photos take longer to decode and use more memory, and a modern browser with WebAssembly plus network access is required.

Common mistakes

• Expecting a skeleton. Only dots are drawn; no lines connect them.
• Reading 553 as a score. It is a maximum count, not a measure of accuracy.
• Uploading a group photo. One person is tracked, and the choice may not be the one you wanted.
• Looking for a download. There is no export button on this page.
• Assuming a low count means failure. Hands out of frame legitimately remove 42 points.
• Using a distant full-body shot. The face and hands need pixels to be found.

Frequently Asked Questions