Bounding Box IoU: Check Overlap and Label Match Separately

Bounding-box intersection over union measures spatial overlap; it does not by itself verify the class label. Record the reference box, coordinate basis, overlap rule and label comparison separately before treating a pair as an accepted annotation match.

This guide is for annotation buyers reviewing dataset deliveries. It does not advertise work opportunities, hiring status or pay. Data Annotation Companies is independent and is not affiliated with CVAT or PyTorch.

What does IoU measure?

The official CVAT automated QA documentation describes box IoU as intersection area divided by union area, and distinguishes shape matching from label matching. Its documented comparison requires both shape and label agreement. The Torchvision box_iou reference likewise describes a geometric overlap result between supplied sets of boxes.

A buyer must still define the reference annotation, task ontology and acceptance policy. A geometric function cannot decide whether the reference itself is correct, whether an object was missed, or whether the assigned category is appropriate. A tool's comparison settings are not a universal contract threshold.

How is the overlap calculated in a simple example?

Illustrative example. Two invented axis-aligned boxes use the same continuous pixel-coordinate basis. Reference box A extends from x = 0 to 10 and y = 0 to 10. Delivered box B extends from x = 2 to 12 and y = 0 to 10. Each has area 100 square pixel units. This example uses width times height without an inclusive-endpoint plus-one adjustment.

Step	Result
Shared width	10 - 2 = 8
Shared height	10 - 0 = 10
Intersection	8 x 10 = 80
Union	100 + 100 - 80 = 120
IoU	80 / 120 = 2/3

The overlap is about 0.667, not 0.8: dividing the intersection by only the reference area uses a different denominator. Retain the unrounded calculation when comparing a boundary value with the agreed rule. These are synthetic coordinates, not an acceptance recommendation or an evaluation result for a provider.

Why does a geometry pass leave the label question open?

Suppose the fictional policy requires IoU at least 0.60 and an exact approved class match. The same pair passes that geometric rule. If A is labelled valve and B is labelled switch, the class comparison fails under that policy. A spatial pass is not a complete pair match.

Conversely, the same class name does not repair poor geometry. Keep both observations and route disputed reference labels or ambiguous objects to adjudication. Do not change the reference or lower a threshold after seeing results simply to make a delivery pass.

Which fields make the acceptance check reproducible?

Field	Evidence
Item	Asset, frame and object IDs
Boxes	Reference and delivered values
Basis	Coordinate format and image version
Rule	Overlap threshold and boundary operator
Labels	Both class IDs and ontology version
Outcome	Geometry, label and reviewed disposition

Keep missing objects, extra objects, duplicate matches and ignored regions in the overall quality policy. A single pair calculation does not define those cases or supply a complete detection metric. Report the relevant denominator and adjudication state when summarizing delivery quality.

Frequently asked questions

Is IoU the percentage of the reference box covered?

No. Its denominator is the union area of the two boxes.

Does IoU of one prove the class label is correct?

No. Identical geometry can still carry the wrong label.

Is 0.60 a recommended universal acceptance threshold?

No. It is a fictional rule used only to explain the example.

Use the bounding-box export coordinate checklist to confirm the representation before comparing boxes. Keep delivery counts and rework in the pilot acceptance register. For provider research, browse the company directory and ask suppliers to demonstrate their actual quality and adjudication methods.