AI Object Detector
100% private — runs on your device, never uploaded. Works offline once loaded.
Upload a photo and this tool finds and labels the objects inside it, drawing a labelled box around each one. Detection runs entirely on your device using a DETR (ResNet-50) model, so your image never leaves your browser.
What this tool does
The AI Object Detector scans a photo and identifies the distinct objects inside it, drawing a coloured bounding box and a confidence-scored label around each one. It then lists how many of each object type it found and lets you download the annotated image as a PNG.
Everything happens locally in your browser. When you click Detect objects, the model file loads once and runs on your device — your image is never uploaded, which makes the tool safe to use on private or sensitive photos.
The model behind it
Detection is powered by DETR (DEtection TRansformer) with a ResNet-50 backbone, running through transformers.js in your browser. DETR treats detection as a direct set-prediction problem, which lets it locate multiple objects in a single pass without hand-tuned region proposals.
The model was trained on the COCO dataset, so it recognises 80 common everyday categories such as people, cars, dogs, chairs, bottles, and laptops. The first run downloads a quantised, one-time model (~40 MB) that is then cached for instant reuse.
Good uses
This tool is handy for quickly tagging what's in a photo, sanity-checking a dataset, building alt-text drafts, teaching how computer vision works, or extracting a count of people or items in a scene.
- Auto-labelling everyday objects in personal photos
- Prototyping and demos where privacy matters
- Counting instances of a category in an image
- Learning how transformer-based detection behaves
On-device quality and caveats
Because the model runs on your device, speed depends on your hardware — a busy image may take a few seconds. Accuracy is good for the 80 COCO categories but limited outside them: it cannot read text, identify people, or recognise brands and rare objects.
A 90% confidence threshold keeps false positives low, which means some genuine but uncertain objects may be missed. Treat the boxes and counts as a helpful estimate rather than a guaranteed inventory.
Frequently asked questions
Is my image uploaded anywhere?
No. The DETR model runs entirely in your browser via WebAssembly. Your photo stays on your device and is never sent to any server.
What objects can it recognise?
It detects the 80 everyday categories from the COCO dataset — people, animals, vehicles, furniture, food, electronics and similar common items. It does not recognise faces, text, logos, or highly specialised objects.
Why does the first run take a while?
The first detection downloads the one-time model (~40 MB). After that it is cached by your browser, so later runs start almost instantly.
Why were some objects missed?
The tool uses a strict 90% confidence threshold to avoid wrong labels. Small, blurry, overlapping, or unusual objects can fall below that bar and be skipped. Clearer, higher-resolution photos improve results.
Can I download the labelled image?
Yes. After detection you get a Download button that saves the annotated canvas — image plus boxes and labels — as a PNG file.
What image formats are supported?
Standard browser-decodable formats like JPG, PNG, and WebP work. Formats your browser cannot decode (such as some HEIC files) may not load; convert them to JPG or PNG first.
Advertisement