Average precision (AP)
Average precision, or AP, summarizes a precision-recall curve as a single number by computing the area under it for one class. It captures how well a model ranks and detects that class across all decision thresholds, and it is the building block of mean average precision.
In this article
Building visual or physical AI?
Let's talk.
What is average precision?
Average precision reduces the precision-recall curve for a single class to one number, roughly the area under that curve. A higher AP means the model maintains high precision even as recall increases, so it both finds most of the real positives and keeps its positive predictions trustworthy across thresholds.
In detection tasks, a prediction is counted as correct only when it overlaps a ground-truth object above an intersection over union threshold, so AP is always reported with respect to a chosen IoU level.
Key takeaways
- AP is the area under the precision-recall curve for one class.
- Higher AP means better precision sustained across recall levels.
- In detection, AP is tied to a specific IoU matching threshold.
How it works
Predictions for a class are ranked by confidence, precision and recall are computed at each threshold, and the resulting curve is integrated to a single value. Different conventions exist for the integration, such as interpolating precision or sampling at fixed recall points, which is why AP definitions can vary slightly between benchmarks.
Before you report one AP number, look at the predictions behind it
Average precision summarizes a detector's precision across every recall level, at one or more IoU thresholds. It's how models on COCO are usually compared, and the single number hides which predictions earned it and which cost it. The quiet problems:
- a confident false positive, which costs more than several low-confidence ones
- in mAP, a rare class with a handful of examples that moves the mean as much as a common one
- missing ground truth boxes that turn correct detections into false positives
In Voxel51, Shared Quickstart has 200 COCO validation images with ground truth object detections and a set of model predictions to compare them against.
Shared Quickstart
200 images
COCO validation images with ground truth object detections and a set of model predictions to compare them against.
Try this: Open a few crowded images and look for predicted boxes with no ground truth box under them. Decide whether the model or the label is wrong.
Explore Quickstart in Voxel51
Free with a work or .edu email.
Why it matters
AP condenses a model's threshold-independent quality on one class into a number you can track and compare. It is the per-class quantity that gets averaged into mean average precision, the headline metric for most object detection and instance segmentation benchmarks.
For how AP and mAP are computed across IoU thresholds, see how object detection models are evaluated.
Frequently asked questions
What is the difference between AP and mAP?
AP is computed for a single class. Mean average precision averages AP across all classes, and often across several IoU thresholds.
Why is AP tied to an IoU threshold?
In detection and segmentation, a prediction only counts as correct if it overlaps the ground truth enough, and IoU sets that bar.
Related terms

Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 11, 2026