Image bounding boxes

When to use it

• Training object detectors
• Counting items in photos
• Locating text regions
• Checking a detector's output

What your file looks like

Each row links one image by URL. The image is shown as it is, and boxes are saved as fractions of its width and height. A file like this is enough:

id,image
1,https://example.com/img/001.jpg
2,https://example.com/img/002.jpg

What reviewers see

The image with a drawing surface, a row of label buttons, and a list of the boxes drawn so far with a remove button for each. It works with mouse, touch and pen.

What a reviewer submits

A box is a fraction of the image: x and y are the top-left corner and w and h the size, all between 0 and 1. Fractions stay correct at any resolution.

{
  "answer": {
    "boxes": [
      {
        "label": "car",
        "x": 0.12,
        "y": 0.4,
        "w": 0.3,
        "h": 0.25
      }
    ]
  }
}

What you get back

Boxes from different reviewers are the same object when their overlap (intersection over union) is at least 0.5 and the label matches. A box survives when a majority drew it, and the exported box is the average of the group.

{
  "n": 3,
  "value": [
    {
      "label": "car",
      "x": 0.13,
      "y": 0.41,
      "w": 0.29,
      "h": 0.24
    }
  ],
  "agreement": null,
  "tie": false
}

Quality checks

• Majority filtering removes stray boxes and averages the rest, which tightens edges.
• Show an example of a good box in the question: tight, whole object, partial objects included or not.
• Review a small sample by hand before the full run.

Tips

• Say what to do with objects that are cut off or hidden. It is the most common source of disagreement.
• Keep images a reasonable size. Reviewers on phones load them over mobile data.
• Use one label per box. If you need several attributes, ask a second question.

Limits

Up to 20 labels and 100 boxes per image. Pixel sizes are not stored, so a COCO or YOLO export needs your own width and height per image.