AI Object Detector

What it does

Upload a photo and the page draws a coloured rectangle around each object it recognises, labels it, and lists the detections with a confidence score. The annotated image can be downloaded as a PNG, and the panel reports how many objects were found.

This is object detection, not classification. A classifier answers "what is this picture of" with one label for the whole frame. A detector answers "where are the things, and what is each one" — which is why it can find three people, a dog and a car in the same scene and give each its own box.

The model

EfficientDet-Lite0 is the small end of the EfficientDet family: fast enough to run on a laptop without a graphics card, and accurate enough for everyday photographs. It knows the COCO vocabulary — people, vehicles, animals, furniture, kitchen items, sports equipment and similar common things. It does not know product models, brands, text or anything trained after that dataset.

What appears on screen

The canvas is sized to the image's natural dimensions and the photo is drawn onto it, so the downloaded file keeps the original resolution rather than the size shown in the panel. On top of that, each detection gets a two-pixel box in a rotating palette, with a filled label strip above it showing the category and a whole-number percentage.

The side panel repeats the same information as a list, with a progress bar per detection. The counter in the header shows the total. Colours are assigned by order, not by class, so the same category can be red in one image and blue in the next.

If nothing crosses the model's confidence threshold, the canvas still shows the photo, the list shows a "no objects detected" message, and the download button stays hidden. There is no annotated file to save in that case.

Step by step

1. Click the upload area and choose an image. The picker accepts any image type; dragging onto the panel is not supported.
2. The first run downloads the MediaPipe bundle, the WebAssembly runtime and the model weights. The panel names each stage.
3. Detection starts automatically once the file is loaded — there is no separate Analyse button.
4. Read the boxes and the list. The percentages are the model's confidence for that label.
5. Press Download to save the annotated canvas as detected-objects.png.

The Load AI Model button warms the model up before you choose a file. When it succeeds it changes to a disabled "Model Ready" state, and subsequent images are processed almost immediately until the page is reloaded.

Reading the scores

A score is the model's confidence that a particular box contains that class. It is not a probability that the label is correct in any absolute sense, and it is not comparable across images. A 60 percent detection in a cluttered photo can be more reliable than an 80 percent detection in an unusual one.

Overlapping boxes are common in busy scenes, where the model is weighing two plausible labels for the same region. Treat the top label as the best guess and the rest of the list as alternatives. A low score across every detection usually means the objects are small, partially hidden, or outside the model's vocabulary.

Where it works well

• People and groups, especially full figures in reasonable light.
• Vehicles in street scenes, car parks and driveways.
• Common animals — dogs, cats, birds, horses — when they occupy a decent part of the frame.
• Household and office objects photographed close up.
• Quick triage of photo libraries: finding images that contain people, or sorting product shots from scenery.

Where it struggles

Small objects are the main weakness. A person in the background of a wide landscape may be missed entirely, and a shelf of products will usually yield a few large boxes rather than one per item. Occlusion hurts too: a dog behind a sofa is harder than a dog on a lawn.

Unusual viewpoints, heavy cropping and extreme lighting all reduce confidence. So does anything the COCO categories do not cover — a specific machine part, a piece of software on a screen, a medical image or a logo. The detector will still draw boxes, but the labels will be the nearest familiar thing.

Dense scenes can produce a pile of overlapping boxes that looks more confident than it is. The ten-result cap also means that a crowd of twenty people will not be fully enumerated.

Uses that make sense

Detection is useful for sorting, counting and cropping. It can tell you which photos in a folder contain people before you publish them, help choose a crop around a subject, or flag images that need a closer look. It is also a quick way to understand what a vision model sees before you build something more specific.

It is not suitable for safety-critical counting, surveillance, accessibility descriptions that must be accurate, or any decision where a missed detection has consequences. There is no tracking between frames, no video mode and no custom training.

Privacy

The image is processed entirely in the browser. It is decoded locally, passed to the WebAssembly model and drawn to a canvas; nothing is uploaded and no results are stored. The page does fetch the runtime and model files from public CDNs, so those servers see that the tool was loaded, but not the picture. The annotated PNG you download stays on your device.

Limits

One image at a time, up to ten detections, no video, no webcam, no batch processing, no export of coordinates as JSON or CSV, no class filtering, no confidence threshold control and no custom classes. Very large images take longer to decode and draw. The page needs a modern browser with WebAssembly and network access on first load.

Common mistakes

• Expecting brand or product names. The labels are generic COCO categories.
• Assuming every instance is found. The cap is ten, and small objects are often missed.
• Reading the percentage as accuracy. It is internal confidence, not a guarantee.
• Using it to read text. There is no OCR; screens and signs become generic objects.
• Forgetting that the download is the annotated canvas. The original photo is unchanged on disk.
• Treating a clean result as proof of absence. A missing box means the model did not see it, not that nothing is there.

Frequently Asked Questions