Spotter
Point your camera, or drop in a photo, and a real AI model boxes and labels every object it sees, live. It recognises 80 everyday things and counts them as they appear, all in your browser. Nothing is uploaded.
🔒 100% private — your camera and photos never leave your deviceSee what AI sees
A neural network finds and labels objects in real time, right on your device. Start your camera, or drop in a photo or video to try it without a webcam.
Loading the AI model…
The detection model (a few MB) downloads once, then is cached for instant, offline use. Everything runs on your device.Spotted so far this session
How it works
Every frame from your camera, or every photo you drop in, is handed to a neural network that has been trained to find objects. For each thing it recognises it returns a box, a label like "person" or "cup" or "laptop", and a confidence score. The page draws those boxes on top of the live image, tallies how many of each kind are in view, and quietly keeps a thumbnail of the first time it sees each new category. All of this happens many times a second, entirely on your machine.
The detector is Google's MediaPipe Object Detector, an EfficientDet model running through WebAssembly and your GPU. Because it runs locally, there is no upload, no round-trip to a server and no privacy trade-off, and it keeps working with your internet switched off once the model has cached.
Allow your camera, or drop in a photo or video to try it with no webcam.
Labelled boxes appear around every object, with a live count of each kind in view.
Raise the confidence threshold to drop weak guesses, or switch to the accurate model.
Save an annotated still as PNG, or record the detection overlay as a clip.
Why run it in the browser?
Object detection is usually something you set up on a server or wire to a paid cloud vision API, sending your images or your webcam off to someone else's machine. Doing it on-device means the footage simply never leaves your computer. It is also instant, because there is no upload to wait on, and free to run, because there is no server bill behind it. It is a neat demonstration of how much real machine learning the modern web platform can do on its own.
Are my camera or photos uploaded anywhere?
No. The AI model and all the drawing run locally in your browser through WebAssembly and your GPU. Your camera frames and any image you drop in never touch a server, which is why it is instant and private.
What can it detect?
It recognises 80 common categories from the COCO dataset: people, cars, bikes, buses, dogs, cats, cups, bottles, laptops, phones, keyboards, chairs, couches, books, and a range of food, among others. It will not detect things outside that set.
How is real-time AI running without a server?
It uses Google's MediaPipe Object Detector, an EfficientDet neural network compiled to WebAssembly that runs on your GPU. The model downloads once, caches, and then analyses each frame on your own device.
Why is the frame rate lower on my phone?
The model is heavier on a phone than on a laptop, so it analyses fewer frames a second. It still works; the boxes just update a little less often. The Fast model is the smoother choice on mobile.
Fast versus Accurate — what is the difference?
Fast (EfficientDet-Lite0) is smaller and runs at a higher frame rate. Accurate (EfficientDet-Lite2) is a larger model that catches more objects and smaller ones, at the cost of speed and a bigger one-time download.
I build tools like this every day.
Senior full-stack engineer, available for senior or contract work, fully remote. See the rest of the lab or get in touch.