LIVE — computed from your camera, on this device

• Object names, confidence and boxes: COCO-SSD (TensorFlow.js), 80 everyday categories.
• Motion regions: difference between consecutive camera frames.
• Tracking: between detections, boxes follow objects by matching small image patches from frame to frame, with predicted motion as a fallback. A box tagged TRK is being followed, not freshly detected.
• Auto-throttle: if the overlay starts dropping frames, detection runs less often until the phone catches up. A ▼ next to the INF time shows when this is active.
• Scanned text and word boxes: Tesseract OCR.
• Scene summary: fixed rules over the detections plus pixel statistics (brightness, colours, sky and greenery).
• Telemetry: resolution, measured video FPS, model time per pass, brightness, zoom, torch state and compute engine.

EFFECT — visual only, no data behind it

• Sweeping scan line, scan lines texture and edge vignette.
• Frame corner brackets and the centre reticle.
• Glow, pulsing dots, bracket snap-in and the spinning lock-on ring.
• The sweep band shown while an analysis runs.

Bracket positions around objects are real. Their animation is an effect.

LIMITS

• Only 80 object categories are known. Anything else is not named.
• People are detected as the generic category “person”. No face recognition, no identification.
• Moving the phone makes the whole frame change. Motion regions are hidden and “camera moving” is shown instead.
• OCR works best on printed text in good light. Handwriting and small text often fail.
• Torch and zoom depend on the phone and browser. Unsupported controls are disabled.
• Voice commands use the browser's speech service, which may send audio to its provider. Everything else stays on the device.