How face detection works

This tool uses the browser's built-in FaceDetector API, available in Chromium-based browsers (Chrome and Edge). The detector runs locally - no image data is transmitted to any server.

Bounding boxes and landmarks

Each detected face is outlined with a colored bounding box drawn on an HTML5 canvas overlaid on the video stream. The native API also exposes landmark points (eyes, nose, mouth) when available.

Facial landmarks: the 468-point mesh

Advanced facial landmark detection (as used in MediaPipe FaceMesh) maps 468 keypoints across the face: the outline of the face, eyebrows, eyes, nose, lips, and iris. Each point has x, y, and z coordinates, enabling 3D face geometry estimation from a single 2D camera. This is the technology behind AR face filters, live makeup try-on, and gaze tracking.

Applications

Augmented reality: virtual glasses, hats, makeup, and masks in real time (Snapchat, Instagram, TikTok filters).
Driver monitoring: detecting eye closure and head tilt for drowsiness alerts; regulatory standards require this in new EU vehicles from 2024.
Accessibility: gaze-based cursor control for users who cannot use hands.
Security: facial recognition and liveness detection for authentication.
Medical facial analysis: detecting facial asymmetry associated with stroke, Bell's palsy, and other conditions.
Virtual try-on: e-commerce tools for glasses, sunglasses, and makeup products.

Accuracy limitations

Face landmark detection performance degrades under several conditions:

Extreme angles: profile views and upward/downward head tilts beyond ~45° reduce accuracy significantly.
Occlusions: face masks, large sunglasses, hats, and scarves covering facial landmarks reduce detection confidence.
Heavy makeup and face paint: theatrical makeup and costume paint can confuse landmark models trained on natural faces.
Lighting: very low light, harsh side lighting, or strong backlighting degrades accuracy.
Multiple faces: performance may drop when many faces are present simultaneously in the frame.

Privacy

All detection runs client-side using your device's on-device API. Your camera feed never leaves your device.