Private browser utility / Social & Creator

Hand Tracking from Photos

Runs entirely in your browser - no upload, no sign-up.

Live workspaceLocal processing

Drop an image hereJPEG, PNG, or WebP up to 20 MB and 16 megapixelsChoose image

01 / Overview

How do you detect hand landmarks in an image?

Choose a photo and detect hands to mark the wrist, finger joints, and fingertips. This image-based hand tracking tool finds up to two hands with 21 points each. Review the overlay, inspect pixel coordinates, then save an annotated PNG or landmark JSON. Your photo stays on your device.

How to use

01
Choose a photo
Drop a JPEG, PNG, or WebP image, or load the sample photo to try the workflow.
02
Detect the hands
Start detection and wait for the local model. The first run needs an internet connection to download model files.
03
Inspect the overlay
Show point numbers, toggle finger connections, and expand the coordinate inspector for either detected hand.
04
Save your result
Download an annotated PNG or copy and download the landmark JSON for a prototype or analysis.

Who it's for

Developers can inspect landmark indexing before building a gesture-controlled prototype.
Artists and animators can study finger positions in reference photos.
Students can compare normalized coordinates with image-pixel positions.
Researchers can check whether a sample image is suitable for a hand-pose experiment.

Use a sharp, well-lit photo with the whole hand visible. The model first locates palms, then estimates 21 landmarks for each detected hand: one wrist and four points along each finger. The overlay connects these points so missing or misplaced estimates are easy to spot.

Detection uses Google MediaPipe Hand Landmarker in a browser worker. The runtime and model download when you start; your photo is processed locally. This page analyzes still images, including saved video frames. It does not access your camera or track motion over time.

The coordinate inspector shows X and Y in source-image pixels. JSON preserves normalized X, Y, and wrist-relative Z. A smaller Z means a point is estimated closer to the camera. This depth is not a measured distance in meters.

FAQ

Is my photo uploaded for hand tracking?

No. Your image is decoded and analyzed inside your browser. Public MediaPipe runtime and model files download on the first run, but the photo and landmark results are not sent to that service. Clear the tool or close the tab to release the current image.

Is hand landmark detection free?

Yes. Detection, the coordinate inspector, annotated PNG downloads, and JSON exports are available without an account, watermark, or payment. Your device performs the work, so speed and available memory determine the practical limits.

What images and limits are supported?

Use JPEG, PNG, or WebP files up to 20 MB and 16 megapixels, with each side at least 32 pixels. The tool finds at most two hands in one image. Clear views with visible fingers work better than tiny, blurry, or heavily covered hands.

Does this tool track hands live with a webcam?

This page detects hand landmarks in still photos. It does not request camera permission or analyze a live video stream. To inspect a particular moment in a video, save that frame as an image and load it here.

What are the 21 hand landmarks?

Point 0 is the wrist. Points 1 through 4 follow the thumb, 5 through 8 the index finger, 9 through 12 the middle finger, 13 through 16 the ring finger, and 17 through 20 the little finger. The last point in each group is the fingertip.

Why can left and right hand labels look wrong?

Mirrored images, unusual views, occlusion, and model errors can affect handedness labels. The image is analyzed as loaded and is not automatically mirrored. The displayed percentage is confidence in that left/right classification, so verify it against the visible hand before relying on the exported label.

Can the model measure hand size or identify gestures?

The output is an estimated landmark arrangement, not a calibrated measurement or gesture label. Normalized coordinates describe positions within the image; relative depth is not a physical distance. Gesture recognition and real-world measurements require additional logic, context, and validation.