What is the difference from object detection?

Detection draws a box around the object; keypoints mark named locations on it, capturing configuration rather than just presence.

What are keypoints used for?

Pose estimation, document alignment, facial landmarks, gesture control and measurement tasks where shape matters.

How is accuracy measured?

By the distance in pixels, often normalized by object size, between predicted and hand-marked points on a labeled test set.

Are keypoints the same as features?

They become features: the coordinates feed downstream models and rules, but they start as detected locations, not raw measurements.

AI Glossary / AI and machine learning / Intermediate

Keypoints

Definition

Keypoints are specific, meaningful locations detected on an object or image, such as the corners of a document, the joints of a body or the landmarks of a face, given as coordinates a model can reason about.

What it means

Some vision questions are not what is in the picture but where its parts are. Keypoints answer with a set of named coordinates: the left wrist, the top-left corner of the form, the center of the logo.

They turn pictures into geometry. Once a person's joints are keypoints, posture, gait and unsafe lifting become measurable; once a document's corners are keypoints, the photo can be deskewed into a flat scan automatically.

Detection is learned from labeled examples: annotators mark the points on thousands of images, and the model learns to predict heatmaps or coordinates for each named point on new images.

Keypoints differ from boxes: a bounding box says roughly where something is, while keypoints describe its configuration, which is what you need when the shape of the thing, not just its presence, carries the meaning.

For a business, keypoints sit inside motion analysis, document processing, augmented reality try-ons and quality inspection. The buyer's question is always the same: which points, labeled how consistently, and with what accuracy on your images.

The same machinery serves faces, hands, products and vehicles: anything with named landmarks whose positions carry the question you actually want answered.

How it works

Define the point set: which named locations matter for the task, written down precisely enough that two annotators mark the same pixel.

Label training images with the points, including occlusion flags for points that exist but are not visible.

Train a detector to predict each point's position, typically as a heatmap per point, and validate localization error in pixels.

Consume the coordinates downstream: angles, distances and alignments computed from points drive the actual business logic.

Monitor accuracy on live imagery, since camera angle, lighting and clothing drift the detector's errors over time.

Three practical examples

A physiotherapy app measuring squat depth from hip, knee and ankle keypoints captured by a phone camera.
A bank's mobile onboarding deskewing photographed ID cards by their four detected corners before reading them.
A warehouse safety system flagging unsafe lifting posture from body keypoints in camera feeds.

Think of it this way

Think of a tailor's pins on a fitting dummy: shoulder, waist, hem, cuff. The pins are not the garment, but once they are placed, every measurement the tailor needs follows from their positions. Keypoints are the pins a model places on an image.

The limit: a pin in the wrong place misleads every measurement built on it. Keypoint errors compound through the geometry, so a two-pixel detection slip can become a five-degree posture error downstream.

Worked example

deskewing a million forms

Suppose an insurer photographs a million paper claims a year, and the documents arrive at every angle. The figures are illustrative.

Corner keypoints on each photo let the pipeline warp every image to a standard flat layout before the text is read.

Read accuracy jumps from 88 to 97 per cent, because the recognizer now sees consistent, squared-up pages.

The four pins did no reading at all; they simply made the reading easy, which is the quiet power of geometry before recognition.

Illustrative case study

A fictional fitness app launched form-checking from body keypoints, tested mostly on gym footage of athletes in fitted clothing.

Real users, baggy shirts, cluttered garages, bad lighting, produced keypoint errors that flipped good-form verdicts to bad, and the reviews turned savage within a week.

The invented lesson: keypoints are only as good as the imagery they were trained on. Label your users' reality, not your demo's.

Common misconceptions and mistakes

Keypoints and bounding boxes are interchangeable. Boxes give location; keypoints give configuration, and posture, alignment and shape questions need the latter.
More keypoints always help. Each point is an annotation cost and an error source; the right set is the smallest that answers the question.
Pixel accuracy is the whole story. Consistency matters as much: a detector that is always three pixels left beats one that averages zero with wild swings, for most geometry.

Business use

Specify the exact point set and annotation rules before buying or building keypoint detection; vague landmarks guarantee vague models.

Validate on imagery that matches your deployment, lighting, clothing, angles, not the vendor's demo set.

Track downstream error, not just pixel error: the business metric lives at the end of the geometry.

Insist on a localization-error figure measured on your own labeled sample before signing off any keypoint-based system.

Check your understanding

Why do keypoints enable measurements that bounding boxes cannot?

Answer: Because keypoints give the positions of named parts, so angles, distances and alignments can be computed from them, while a box only says roughly where the whole object sits.

Related terms

Learn next

AI for Business Owners covers how to scope a vision vendor's claims down to the specific measurements you are buying: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Key Value (KV) store Next: Keyspace (distributed data store)