See what the model sees.

Real-time object detection on your webcam or any image, running entirely in this tab. Nothing is uploaded.

Recorded example. Pick a live source to load YOLO11n (~11 MB).

Confidence 35%

Tracked class Person IDs persist across frames

Playback rate 23.98_fps Video frame rate, not inference speed

Model YOLO11n + ByteTrack, pre-rendered

Execution Playback No frames leave this device

Detections

RECORDED

Boxes and IDs in the recording are baked into the video. Start the webcam or analyze an image to inspect live confidence scores here.

What happens on each frame

01Capture
A webcam frame or your image is drawn to a canvas, capped at 1920 px.
02Letterbox
Scaled into a 640 × 640 grey-padded square and packed into an RGB float tensor.
03Infer
YOLO11n runs through ONNX Runtime Web on WebAssembly, on your CPU.
04Decode
Top class per anchor, mapped back to image pixels, then class-aware NMS at IoU 0.45.

The recorded clip is pre-rendered output from the same YOLO11n weights plus ByteTrack, tracking people only. Its 23.98 fps is video playback, not inference speed, and live mode shows detections without persistent track IDs. The webcam view is mirrored so it reads like a mirror; exported coordinates are unmirrored. Weights: webnn/yolo11n, an ONNX export of Ultralytics YOLO11n (AGPL-3.0). Footage: Yonge–Dundas crossing, Toronto, by Raysonho (Wikimedia Commons, CC0).