SYSSEDER / BROWSER AI LAB
MediaPipe Hand Landmark Playground
Find up to two hands in an image, then inspect the model's 21-point skeleton. No camera or model starts automatically.
Downloads Google Hand Landmarker float16 v1: 7,819,105 bytes (7.46 MiB) from storage.googleapis.com, plus MediaPipe Tasks Vision 0.10.21 JavaScript and WebAssembly from cdn.jsdelivr.net (about 9.5 MiB additional uncompressed files; transfer sizes and cache use vary). No API key or paid service.
Model not loaded.
Delegate: not initialized.JPEG, PNG or WebP only; at most 10 MiB and 16 million decoded pixels. Images are scaled to fit 1280 × 1280 before inference. Camera preview is unmirrored.
No inference yet.
Read results carefully: the handedness score concerns the left/right classification, not landmark accuracy or certainty that a person has been identified. A drawing may produce zero hands. This is not identity recognition, gesture interpretation, a medical assessment or a measurement instrument.
Privacy, versions and licenses
This playground's code processes selected images and camera frames in browser memory; it does not upload, store or log them. Loading fetches model/runtime files from Google and jsDelivr. This Blogger page can also run advertising, analytics and other third-party scripts: this is not a claim that the entire blog is offline or private. Do not use sensitive images. Stop clears the canvas and releases camera tracks; leaving the page also stops capture. Browser caches may retain downloaded model files.
MediaPipe Tasks Vision 0.10.21 and Google's model card identify Apache-2.0 licensing. Read the library license, model card and limitations, and official web guide. Model: hand_landmarker/hand_landmarker/float16/1/hand_landmarker.task.