AI Hand Tracker

What it draws

Upload a photo containing one or two hands and the page places a small ring at each detected joint: the wrist, four points along each finger, and a point at each fingertip. The panel reports the total count, up to 21 per hand. The overlay sits on the photo at full resolution, and the model runs entirely in the browser.

There is no download button on this page. The tracking result is for inspection on screen; keeping it means taking a screenshot. There is also no gesture name, no handedness label and no export of coordinates.

The model

The 21-point scheme is compact and standard: a wrist anchor, then four joints and a tip for each of the five digits. That is enough to describe an open palm, a fist, a point or a pinch, and not enough to describe skin creases, nail position or the exact curvature of a finger. The model predicts the points it cannot see, which is why a hand partly behind an object still produces a complete set, sometimes in the wrong place.

Two hands are tracked, but the page draws both in the same colours. The underlying result includes a handedness classification for each hand, yet the interface does not display it, so you cannot tell from the overlay which hand the model considers left or right.

What the overlay looks like

Each landmark is drawn twice: a filled white circle of radius three, then a blue ring of the same radius stroked over it. The result is a clean dot that stays visible on both light and dark images. No lines connect the points, so the hand reads as a constellation rather than a skeleton — useful for seeing raw joint positions, less useful for instantly reading a pose.

The canvas is sized to the image's natural dimensions, so the dots land on the correct pixels regardless of how large the photo is. The count badge shows the number of hands multiplied by 21, which means a single detected hand reports 21 even if some points are estimates rather than confident detections.

Step by step

1. Click the upload area and pick a photo. Drag and drop is not supported.
2. On first use, the browser downloads the runtime and the hand model; the panel reports each stage.
3. Inference starts automatically after the image loads.
4. Check the dots. Fingers that curl away from the camera usually produce points that bunch together.
5. Upload another image to run again; the loaded model makes repeat runs near-instant.

The Load AI Model button preloads everything ahead of time. On success it turns green and reads "Model Ready", and further images are processed without the initial delay.

Where it is useful

• Animation reference. Trace a photographed hand pose to plan a drawing or a 3D rig.
• Gesture prototyping. Check whether a camera angle captures enough of the hand before building a live gesture interface.
• Ergonomics. See how a hand sits on a mouse, a tool or an instrument.
• Teaching anatomy. A visual demonstration of where the joints of the hand fall in an image.
• Sign-language research. Landmark data is a common input for handshape analysis, though this page only displays it.

It is not a gesture recogniser. It knows where the joints are, not what the hand is doing. A thumbs-up, a peace sign and an open palm all produce 21 points; the meaning is left to you.

Where it fails

Self-occlusion is the main problem. When fingers overlap, curl toward the palm or point at the camera, the model has to infer depth from a flat image and often places points on top of each other. A fist can look like a cluster rather than five folded digits.

Partial hands are another weakness. A hand cut off by the frame edge, hidden behind a cup or covered by a sleeve gives the model less evidence, and it may still output all 21 points with low confidence — the interface has no confidence readout, so there is no warning that the fit is poor.

Small hands, motion blur, low light and heavy shadows all reduce accuracy. So do unusual skin tones against similar backgrounds, though the model is generally robust here. Very large images are fine, but a hand occupying a tiny part of a wide shot is effectively a small hand.

Two hands close together can also be merged or swapped, particularly when they touch. The page draws both sets without labels, so a swap is invisible in the overlay.

What it does not do

No gesture names, no left/right labels, no confidence scores, no connections, no download, no coordinate export, no video, no webcam, no batch processing and no multi-frame tracking. It does not measure finger length, joint angles or hand span, and it does not build a 3D model you can rotate.

It also does not identify people. Hand geometry is less distinctive than a face, but a photo of a hand is still personal data, and the same consent questions apply when the hand belongs to someone else.

Privacy

The image is decoded locally, passed to the WebAssembly model and drawn to a canvas. Nothing is uploaded and no result is stored after the page is closed. The runtime and model weights are fetched from public CDNs on first use, which reveals that the tool was opened but not what was in the photo.

Limits

One image per run, at most two hands, 21 points each, no download and no numeric output. There is no threshold control, no way to select which hand to track and no way to hide one hand's points. Large photos take longer to decode, and the page requires a modern browser with WebAssembly plus network access on first load.

Common mistakes

• Expecting a gesture label. Only joint positions are shown.
• Looking for a download. There is none; the overlay is on screen only.
• Assuming 21 dots means a good fit. The model always returns a full set, even when guessing.
• Uploading a hand at the frame edge. Partial hands produce unreliable points.
• Reading the count as hand count. The badge multiplies hands by 21.
• Using a low-resolution crop. Small hands lose the detail needed for stable joints.

Frequently Asked Questions