This free app finds and tracks objects in your webcam feed or in a video file, live. The AI model (YOLO26 or YOLOE) runs on your own graphics card through WebGPU and ONNX Runtime Web, so there is no server, no account and no install, and your video never leaves your device.
car #12) and its box glides with it between frames.| Mode | Model | Objects | Accuracy (mAP) | Model FPS | Download |
|---|---|---|---|---|---|
| Fast | YOLO26n | 80 | 0.380 | 7–8 | 5 MB |
| Accurate | YOLO26s | 80 | 0.475 | ~2 | 19 MB |
| More objects | YOLOE-26s | 268 | 0.370 | ~2 | 21 MB |
Accuracy on 500 COCO val2017 images; speed on a laptop with Intel UHD 620 graphics (Chrome, WebGPU).
Yes. It is free and open source (Apache-2.0). There is no account, no install and no usage limit, because everything runs on your own device.
No. Video frames are processed by the AI model inside your browser, on your own GPU, and are never sent to a server. The site only counts anonymous page views.
It works best in Chrome or Edge 113 or newer on a laptop or desktop, using WebGPU. Firefox and Safari use a slower WebAssembly fallback. Camera access requires HTTPS, which this site uses.
On 500 COCO val2017 test images, the Accurate mode (YOLO26s) scores 0.475 mAP and the Fast mode (YOLO26n) 0.380 mAP. On a laptop with integrated graphics, Fast processes 7 to 8 frames per second, and boxes are redrawn at about 60 FPS in between.
Yes, in the More objects mode. It uses an open-vocabulary YOLOE model with 268 everyday object classes, including smartwatch, glasses, earbuds, keys and wallet. The Fast and Accurate modes detect the 80 standard COCO classes.
Yes. Click Upload video to run detection on a video file from your device. Each car or person gets a stable tracking ID and its box follows it as it moves. The video never leaves your device.