Learn AI /Computer Vision and Language Technology /Pose estimation
Pose estimation
- Deeper
- Computer Vision and Language Technology
- Technical readers
Also called: keypoint detection.
Back to the milestone you came from
In five seconds
What it is
A computer vision method that locates important body, face or hand points in an image or video and often links them into a skeleton-like layout.
In plain words
It finds important points on a body, face or hand in an image, such as elbows or eyes, so software can describe posture or movement.
An everyday example
A basketball coach prints a photo of a player taking a shot. On the paper, the coach marks dots on the player's shoulder, elbow, wrist, hip, knee and ankle, then draws simple lines between them to check the shooting form. The coach is not writing a full description of the whole player or colouring every part of the body. The goal is to capture the arrangement of key joints so the posture can be analysed.
A human coach can use wider context, such as knowing the sport, the player and what happened before the photo. A pose estimation model usually relies on visual patterns in the given image or frame and may fail if joints are hidden, blurred or outside the training examples. Also, the model does not literally identify bones or physical mechanics; it predicts point locations.
Remember thisPose estimation predicts where predefined body, face or hand landmarks are and how they connect. The points are estimates with confidence scores, not a definitive account of a person’s movement or intent.
Understand it
Understand it
Why it matters
Pose estimation matters when the arrangement of parts is more important than simply detecting an object. It can support movement analysis, sign and gesture interfaces, sports review, ergonomics, animation, retail analytics and some accessibility tools. For decision-makers, the key question is fit for purpose: do you need a box around a person, or do you need estimated body landmarks? It also matters for risk. Errors are not random. They can increase with occlusion, camera angle, clothing, skin-tone and lighting conditions, crowding, disability-related movement patterns, or underrepresented body types in training data. That affects fairness, safety and whether human oversight is required.
How pose estimation turns an image into keypoints
After reading this diagram, the reader should be able to explain that pose estimation outputs coordinates for body parts, not just a label or a box.
01
1. Input image
Start with a photo or video frame that contains one or more people, faces or hands.
Show a simple scene with two human figures in a rectangular image frame. Keep details minimal and high contrast.
02
2. Target region found
Many systems first locate the relevant subject, often with a bounding box around each person or body part of interest.
Overlay one or two dashed bounding boxes on the figures. Add a small note: 'possible first stage'.
03
3. Keypoints predicted
The model estimates coordinates for landmarks such as shoulders, elbows, wrists, hips, knees and ankles, each with an optional confidence value.
Replace or supplement boxes with coloured dots at joints. Add tiny confidence labels like 0.94, 0.88 next to a few points.
04
4. Pose output
The final output is a set of point locations, sometimes connected into a skeleton for easier viewing or further analysis.
Connect the dots with thin lines. Beside the figure, show a compact table with rows like 'left_wrist: x=214, y=168, conf=0.88'.
A left-to-right flow with four panels. Read from the input image, to person region selection, to keypoint prediction, to structured output. Each panel adds one layer of information so the reader can see how the result becomes more detailed than a class label or bounding box.
Hovering over each stage highlights what new information has been added: image only, box, keypoints, then structured coordinates and links.
Describe this without the diagram
The diagram is a four-step flow read from left to right. First, there is a plain image frame with two people. Second, dashed boxes appear around the people to show that some systems first locate a target region. Third, coloured dots appear on major joints such as shoulders, elbows, wrists, hips, knees and ankles. Fourth, thin lines connect these dots, and a small table lists coordinates and confidence values. The message is that pose estimation produces structured landmark locations, which is more specific than saying a person is present and different from marking every pixel.
With reduced motion: Show the same four panels side by side without animation. Use static arrows between panels and a legend explaining boxes, keypoints, lines and confidence values.
Pose estimation takes an image or video frame as input and tries to find specific points of interest, often called keypoints. For a person, these points might include the nose, shoulders, elbows, wrists, hips, knees and ankles. The output is usually a list of coordinates, sometimes with a confidence score for each point.
In many systems, the model first finds a person or another target object in the image. Then it predicts where each keypoint probably sits. Some systems also connect the points with lines to show body structure, but those lines are a visual aid rather than a real bone map.
This is different from image classification, which gives one label for a whole image, such as "person running". It is also different from object detection, which draws a box around a person, and from image segmentation, which marks many pixels belonging to that person. Pose estimation goes further by estimating the arrangement of parts.
The method is useful because position alone is often not enough. A bounding box can tell you that a person is present, but not whether their arm is raised, whether they are bending, or how a hand is shaped. That said, the output is still an estimate from visual patterns. It can be wrong when body parts are hidden, blurred, crowded, unusually positioned, or poorly represented in the training data.
How “A sports coach marking joints on a training photo” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Dots marked on joints | Predicted keypoints with image coordinates |
| Lines between the dots | Skeleton-style links showing body structure |
| Checking shooting form from joint arrangement | Using pose to infer posture or movement |
| Not colouring the full body | Different from image segmentation |
A worked example
A video producer checks one frame from a consented, staged recording of a presenter demonstrating an arm movement.
Starting from: One 1280 by 720 pixel frame containing one presenter standing sideways to the camera.
- A top-down pipeline first detects the presenter and crops the relevant region of the frame.
- Within that crop, the model predicts predefined body keypoints, including the left shoulder at x=640, y=250 and the left wrist at x=780, y=190.
- It returns confidence scores for each point, such as 0.97 for the shoulder and 0.91 for the wrist. A partly hidden ankle receives 0.62.
- The system links compatible keypoints into a skeleton-like layout and derives an elbow angle of about 156 degrees.
Result: The output consists of estimated landmark coordinates, confidence scores and a derived angle. It can direct the producer to frames needing review, but it does not establish whether the movement was performed correctly.
In a United Nations workplace
Before clearing an implementing partner’s progress report for the front office at close of business, a partnerships officer in a documentation and library service checks an approved sample of digitised instructional videos. The partner reports that presenters remain fully framed throughout each recording.
Pose estimation can locate shoulder, elbow, wrist and hand landmarks frame by frame, with confidence scores, to identify passages where a presenter’s hands or upper body may fall outside the image. This can support a targeted quality check, but it cannot determine whether the instruction itself is accurate or accessible.
Before you try thisDo not submit identifiable footage to a tool that has not been approved for personal data. Confirm the permitted purpose, access controls and retention arrangements before processing. Low confidence, occlusion and framing changes can produce false flags, so the officer should review the relevant passages rather than treating the model’s output as evidence of partner performance.
Where you may hear it
- A vendor demonstration for an approved video-processing tool
- A procurement meeting on movement-based learning software
- An accessibility working session on signed video content
- An evaluation office review of a computer-vision methodology
Avoid confusion
Avoid confusion
Often confused with
Object detection
A computer vision task that finds instances of chosen object types in an image or video and marks where each one appears.
Image segmentation
Image segmentation divides an image into pixel-level regions; semantic segmentation labels them as roads, buildings or water.
Facial recognition
A computer vision method that compares face images or representations to estimate whether they show the same person.
Common misconceptions
Common misconception: Pose estimation and object detection are basically the same thing.
Correction: They are related but not the same. Object detection usually answers 'where is the person?' with a bounding box. Pose estimation answers 'where are the important body parts?' with landmark coordinates. A system may use detection first, but the outputs and use cases differ.
Common misconception: If the skeleton overlay looks neat, the measurements must be accurate.
Correction: A tidy visual can hide uncertainty. Many systems smooth outputs over time or draw standard links between points, which can make errors look plausible. Confidence values, camera angle, occlusion and validation on the relevant population still matter.
Common misconception: Pose estimation recognises who a person is.
Correction: Not by itself. It estimates body, face or hand landmarks. Facial recognition tries to match identity across images. A pose model may process a face or body without identifying the person, though privacy risks can still arise from the surrounding system.
Quick check
Which output best matches pose estimation for one person in a photo?
Pose estimation produces structured landmarks, usually as coordinates for predefined body parts. A label is image classification, a box is object detection, and a paragraph is image captioning or another language output.
Go deeper
Go deeper
The simple version
It finds important points on a body, face or hand in an image, such as elbows or eyes, so software can describe posture or movement.
The technical version
Pose estimation predicts structured landmarks from visual input. In human pose estimation, a model maps an image or video frame to 2D or 3D keypoint locations for predefined joints or landmarks, often with per-keypoint confidence scores and sometimes part-association links. Common approaches include top-down pipelines, which detect each person first and then estimate keypoints within each crop, and bottom-up pipelines, which predict keypoints for the whole image and then group them into individuals. The result may be represented as coordinates, heatmaps, offset vectors or graphs. It differs from detection, which outputs object-level localisation, and from segmentation, which outputs pixel-level masks. Performance depends heavily on annotation quality, viewpoint, scale, occlusion handling, temporal consistency in video, and how well the training data covers the target population and context.
Reviewed external resources
EncyclopaediaArticulated body pose estimation (opens in a new tab)Wikipedia
This link gives a broad overview of articulated body pose estimation and suits readers who want background, terminology and related concepts in one place.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This link gives short glossary-style definitions of machine learning terms and suits readers who want quick clarification of vocabulary used around pose estimation.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This link gives beginner-friendly machine learning foundations and suits readers who want the wider context behind how vision models are trained and evaluated.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Pose estimation (opens in a new tab)YouTube search
This link is a YouTube search results page, so it gives a range of video explanations with different levels of depth and suits readers who prefer visual teaching and comparison across presenters.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Pose estimation (opens in a new tab)Reddit search
This link is a Reddit search results page, so it gives community questions, practical discussions and varied opinions and suits readers who want informal troubleshooting themes rather than a single authoritative explanation.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Pose estimation predicts where predefined body, face or hand landmarks are and how they connect. The points are estimates with confidence scores, not a definitive account of a person’s movement or intent.