What is bounding box?

A bounding box is a rectangle drawn around an object in an image or video frame, with a label saying what the object is. It is the most common computer vision annotation, used to train models that detect objects, from pedestrians for self-driving cars to products on a shop shelf.

What the work looks like

You open an image in an annotation tool, draw a tight box around every object of the requested classes, and pick a label for each. Guidelines set the details: how tight the box must be, whether to include parts hidden behind other objects, the smallest size to annotate, and what to do with reflections or pictures of objects.

Quality rules that trip people up

• Boxes should touch the visible edges of the object, with no padding.
• Every instance counts, including partly visible ones, unless the guidelines say otherwise.
• Consistency matters more than speed: reviewers compare your boxes with gold tasks.

Example

A street photo contains four cars, one cyclist and a car on a billboard. The annotator boxes the four cars and the cyclist, and skips the billboard car because the guidelines exclude pictures of objects. For finer outlines, see semantic segmentation.