Bounding Box Annotation
What is bounding box annotation?
Annotation with a bounding box means that a rectangle (very rarely an actual square) has to enclose each object in an image or a frame of video. The annotation produces a pair of x min/y min and x max/y max values.
Where can it be used?
Bounding box annotation is one of the most prevalent types of labeling in computer vision projects, given its low cost. As almost anything can be labeled, measured and counted with a bounding box, they are common in every industry.
- Medical: detecting abnormal cells in blood smears
- Geospatial: providing a headcount of cattle in a field by using drones
- Automotive: detecting pedestrians and vehicles for self-driving cars
- Industrial: counting the number of manufactured products
- Agriculture: estimating plant size and count
- Retail: tagging of existing products on supermarket shelves
- Bounding boxes are quick and easy to draw.
- They can be used to extrapolate an object’s actual size even if that object is partially occluded.
- Can easily include a lot of pixels that don’t belong to the object (e.g. background, other objects in front of or around the target one).
- Tools may vary in how fast the annotation can be performed (e.g. two clicks, click and hold, switch to edit mode, start a new box).
- When creating the annotation instructions, make sure to address how tight boxes should be (sometimes slightly looser might be better) and how annotators should handle occlusion, truncation, very small and blurry objects, etc.
- For more precise annotation of rotated objects, some tools support a “rotated bounding box” format in which a degree of rotation is added to the x and y coordinates.
- Some tools provide a crosshair visualization which shows the x and y axis of where the cursor is placed. This can significantly improve the accuracy of the annotation and having to spend additional time adjusting the created box.