Loading tool…
Object detection answers two questions at once: what is in a picture, and where. Drop in photos to get labelled boxes and per-class counts, or switch to the live camera to watch detection happen in real time. Photos and video stay on your device.
Processing locally on your device
Overview
Image classification gives a whole picture one label; object detection finds every instance of known classes and returns a box, a label and a confidence score for each. Both models here are trained on the COCO dataset, which covers 80 everyday categories: people, vehicles, common animals, furniture, kitchenware, electronics, sports equipment and more. Anything outside those categories, such as a specific product or plant species, won't be detected.
Photos are analysed with DETR (DEtection TRansformer) with a ResNet-50 backbone, an Apache-2.0 model from Facebook AI Research, run through Transformers.js. Rather than proposing thousands of candidate boxes and pruning overlaps, DETR predicts a fixed set of objects directly, which makes it reliable in crowded scenes. The 8-bit version used on the CPU is a 43 MB download and takes about 3–4 seconds per photo on a laptop; with WebGPU a 16-bit version (84 MB) runs on your graphics card. Boxes are mapped back to the photo's original resolution.
The live camera uses a different model built for speed: EfficientDet-Lite0 through Google's MediaPipe, a 7 MB Apache-2.0 model that runs on the page with WebGL where available. It processes each new video frame while the tab is visible, typically at real-time frame rates on laptops and recent phones. Detection results are drawn over the video, counted per class, and can be saved as a snapshot PNG.
Every detection is kept down to 25% confidence, so moving the minimum-confidence slider updates the boxes and counts instantly without re-running the model. Class chips let you hide categories you don't care about. Exports respect the current filter: an annotated PNG at full resolution for each photo, a ZIP for a batch, and a JSON file with labels, scores and pixel boxes.
| Photos (DETR ResNet-50) | Live camera (EfficientDet-Lite0) | |
|---|---|---|
| Download | 43 MB (CPU) / 84 MB (WebGPU) | 7 MB |
| Speed | About 3–4 s per photo on a laptop CPU | Real time on most laptops |
| Strengths | Crowded scenes, overlapping objects | Speed, low power |
| Runs on | Web Worker (WebGPU or WebAssembly) | Page (WebGL or CPU) |
| Classes | 80 COCO | 80 COCO |
Drop one or more photos, or switch to Live camera and press Start camera.
Wait for the model to download once; detection starts automatically.
Move the minimum-confidence slider until the boxes look right.
Click class chips to hide categories, and read the counts table.
Download an annotated PNG, a ZIP of all annotated photos, or JSON with every box.
Use cases
Get a quick count of people in a room, cars in a car park or cups on a table from a single photo, then check the boxes visually before relying on the number.
Generate draft boxes for a batch of photos and export JSON with pixel coordinates, then correct them in your labelling tool instead of drawing every box by hand.
See how confidence thresholds and crowded scenes affect detections, and compare a transformer detector on photos with a lightweight real-time detector on the camera, with no Python setup.
Check what appears in photos from a home, office or workshop without uploading them anywhere, because both models run locally in the browser.
Pro tips
Answers to common questions about the object detector.
The 80 COCO categories, including person, bicycle, car, bus, dog, cat, bird, chair, couch, bottle, cup, laptop, phone, book and sports ball. Objects outside that list are not recognized.
They come from different models. Photos use DETR, which is slower but more accurate, especially with overlapping objects. The camera uses EfficientDet-Lite0, which is small enough to run on every frame in real time.
Usually because they are very small in the frame, partly hidden, seen from an unusual angle, or not one of the 80 classes. Cropping, a lower confidence threshold or a better-lit photo help with the first three.
For each photo: the file name, size, counts per class, and every visible object with its label, score and box as [x, y, width, height] in the original image's pixels. Only objects above the current threshold and not hidden by the class filter are included.
No. Frames are analysed in memory and discarded. Nothing is stored unless you press Snapshot, which downloads one PNG to your device.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The object detector is free to use with no sign-up, no watermarks and no daily limits.
Find duplicate and similar images.
Describe images for accessibility.
Extract text from images.
Search your photos by description.
Remove image backgrounds in one click.
Upscale photos 2× or 4× with AI.