Object Detector

Drop, paste or click to choose an image

Min. confidence 50%70%90%Detect objects

Ready. The model downloads the first time you run it (~170 MB).

Object detection goes a step beyond image recognition: rather than saying what a picture is about, it locates each item it recognises and marks it with a box and a label such as person, car, cup or dog, along with a confidence score. This tool lets you try that on your own images in a few clicks.

All processing happens in your browser. The photo is not uploaded, which makes the tool suitable for pictures of your home, workplace or family that you would rather not share with an online service.

How to use it

1. Choose or drop an image file onto the tool.
2. The first time, wait for the model to download; it is about 170 MB and is cached for future visits.
3. Click detect. After a few seconds, coloured boxes with labels and confidence percentages appear over the image.
4. Adjust the confidence threshold to hide uncertain detections, then save or screenshot the annotated result.

How it works

The detector is DETR with a ResNet-50 backbone (Xenova/detr-resnet-50), run through Transformers.js. ResNet-50 is a convolutional network that converts the photo into a map of visual features. A transformer then looks at the whole map at once and proposes a fixed set of possible objects, each with a box position and a class. The model was trained on the COCO dataset, so it recognises about 80 everyday categories, and the tool draws only the predictions above your chosen confidence level.

Good for

• Counting people, cars or animals in a photo as a rough estimate.
• Teaching or learning how computer vision models see and label scenes.
• Checking whether a product photo or scene contains the items you expect.
• Creating annotated example images for presentations or tutorials.

Limitations

• It only knows the roughly 80 COCO categories; anything outside that list is ignored or given the closest wrong label.
• Small, partly hidden or unusually shaped objects are often missed, and crowded scenes can produce overlapping or duplicate boxes.
• Being a compact model running on your device, it is less accurate and slower than large cloud vision systems; desktop Chrome or Edge is recommended.

FAQ

More tools