Computer vision / By James Sun · SurfAI
Motion & gesture lab
Your camera, a model, and a clearer view of the prediction.
Put an idea in motion.
Start your camera to explore gestures and pose landmarks. Your video stays on this device.
Start the camera, then hold one supported gesture in view.
Model confidence is not measured accuracy. Uncertain or unsupported poses are shown as uncertain rather than guessed.
How does recognition work?
The hand mode uses MediaPipe’s trained gesture classifier with landmark tracking. The interface requires a consistent, high-confidence result across several frames before naming a gesture. Body mode uses pose landmarks and geometric rules for raised arms; it does not diagnose exercise form.
Why do camera FPS and model FPS differ?
The camera renders independently while a worker processes smaller frames at an adaptive rate. That keeps the preview smooth without running a large model on every video frame. Available resolution and frame rate depend on your device, browser, camera, and lighting.
Camera permission and privacy
Use HTTPS or localhost. Permission is requested only when you press Start. Video frames are processed in your browser, never uploaded. Stop releases the camera and model resources; leaving the tab pauses the session. MediaPipe code and model files are fetched from their public CDNs.