About Object Detection
Object Detection finds and labels the things in a photo: people, cars, dogs, laptops, cups and dozens more. Each object gets a box, a name and a confidence score, and you can export everything as an annotated image or as structured JSON.
How it works
The tool uses DETR (DEtection TRansformer), a model published by Meta AI. A ResNet-50 network first extracts visual features from the image. A transformer then looks at the whole scene at once and predicts a set of objects directly, each with a class and a box. This approach handles crowded scenes well because the model reasons about objects in relation to each other.
The model was trained on the COCO dataset, a standard benchmark with 80 everyday categories. They cover people and animals, vehicles, furniture, electronics, kitchen items, food and sports equipment.
What you can use it for
- Counting people, cars or items in a photo without doing it by hand.
- Prototyping computer-vision apps: test what an off-the-shelf model sees before you build anything.
- Dataset preparation: export bounding boxes as JSON as a starting point for labeling.
- Accessibility and alt text: get a quick list of what's in an image.
- Teaching and demos about how modern AI vision works.
Tips for the best results
- Lower the confidence threshold to reveal smaller or partly hidden objects. Raise it to remove false alarms.
- Hover over an item in the results list to highlight its box when the scene is busy.
- Objects outside the 80 categories (for example tools or specific brands) won't be labeled, or may get the nearest category.
- The JSON coordinates are in pixels of the original image, ready to use in your own code.