AI Object Detector

Detect and label objects in images using AI — powered by DETR ResNet-50, runs 100% in your browser.

All processing runs entirely in your browser using WebAssembly. No image or text is ever uploaded to a server. First run downloads the AI model (~44–70 MB) and caches it locally — subsequent runs are instant.

Confidence Threshold: 40%

Try sample:

Drop an image here, or click to browse

Supports PNG, JPG, WebP, SVG images

AI Object Detector — Identify Objects in Photos

This tool uses DETR (Detection Transformer) with a ResNet-50 backbone to identify and locate objects in images entirely within your browser. DETR treats object detection as a direct set-prediction problem using a transformer encoder-decoder architecture, eliminating the need for anchor boxes. The quantised model (~160 MB download, cached locally after the first run) can detect 91 COCO classes including people, animals, vehicles, furniture, and everyday objects. Detected objects are highlighted with colour-coded bounding boxes and confidence scores. No image leaves your device.

Built and maintained by Meet Shah · Last updated 19 July 2026

What this tool is used for

Labelling objects in an image without sending it anywhere.
Checking what a detection model finds in a photograph.
Producing bounding boxes for a quick annotation task.
Demonstrating in-browser inference for a talk.
Testing how a model handles a difficult image.

Frequently Asked Questions

What model does the detection use?: DETR with a ResNet-50 backbone, running in your browser through Transformers.js. DETR treats detection as a set-prediction problem with a transformer rather than as region proposals, which is why it needs no non-maximum-suppression step to clean up overlapping boxes.
Why is the first run slow?: Because the model weights have to download and compile before anything can be detected — tens of megabytes, once. After that it is cached, and subsequent images are fast. That download is the entire cost of running inference locally instead of sending your image to a server.
What does the confidence score mean?: The model's own estimate that the box contains that class, from 0 to 1. It is not a probability in any calibrated sense — models are routinely overconfident — so treat it as a ranking rather than a measurement, and set the threshold by looking at results rather than by picking a round number.
Which objects can it recognise?: The 80 COCO categories it was trained on: people, vehicles, animals, common household and food items. Anything outside that list cannot be returned at all — it will either be ignored or labelled as the nearest class it does know, which is worth remembering before reading a wrong label as a mistake.
Is my image uploaded anywhere?: No. The model runs in your own tab via WebAssembly and WebGPU, so the image is never transmitted. That is the practical argument for local inference: it works on photographs you would not send to an API, and it keeps working with no network at all once the weights are cached.

Common errors and gotchas

Treating detections as reliable, since confidence scores are not calibrated probabilities.
Expecting consistent results across runs, when preprocessing and thresholds shift the labels.
Using it on an image class the model was not trained for, where it labels confidently and wrongly.
Reading a missing detection as absence rather than as a limit of the model.
Running it on a very large image in a browser tab, which can exhaust memory.