A 2D bounding box, often shortened to bbox, is a rectangle drawn around an object in a flat image to mark its location and approximate size, without capturing depth or orientation the way a 3D bounding box does. It's one of the most common and basic forms of object labeling in computer vision.

‍

‍Why it matters

Bounding boxes are popular specifically because they're fast to create and easy for a model to learn from, compared to a more detailed method like segmentation which requires tracing an object's exact outline. That speed comes with a real tradeoff though, a box around an oddly shaped object, like a person with an outstretched arm, ends up including a lot of background space that isn't actually part of the object, which is fine for some tasks and a real limitation for others that need precise boundaries.

‍

‍When it's the right choice

Bounding boxes work well for tasks like counting objects, detecting whether something is present, or general object tracking, where knowing roughly where something is matters more than its exact shape. Tools like CVAT and Labelbox make this the fastest annotation type to produce, which is why teams default to bounding boxes whenever the task doesn't specifically require the extra precision that segmentation or polyline annotation provides.