Pose Estimation
Finding the positions of body joints (shoulders, elbows, knees...) in an image or video, producing a stick-figure skeleton for each person.
Also keypoint detection · human pose estimation · skeleton tracking · OpenPose
The computer looks at a photo of someone dancing and puts a dot on each joint, then connects the dots into a stick figure that moves just like the dancer.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Uses / built on
How it works
Object detection draws a box around each person. Pose estimation goes inside the box and finds the body’s keypoints: nose, eyes, shoulders, elbows, wrists, hips, knees, ankles. Connected, they form a skeleton.
Two strategies:
- Top-down: detect each person first, then find the keypoints within each box. Accurate, but slower with many people.
- Bottom-up: find every keypoint in the image at once, then figure out which ones belong together. OpenPose (2017 to 2018) made this work in real time.
Most models predict a heatmap per keypoint, a map of how likely each pixel is to be, say, the left wrist, using CNNs or Vision Transformers.
Applications range from fitness and physical therapy apps to sports analytics, sign language recognition, augmented reality filters and motion capture for animation. A skeleton is also far less identifying than the full video, which can help with privacy.
Everyday use cases
- Case 01
Fitness apps
Counting squats and checking your form through the phone camera.
- Case 02
Motion capture without suits
Animating game characters or video-call avatars from ordinary video.
Strengths, weaknesses and when to use it
Strengths
- Turns video of people into compact, privacy-friendlier skeleton data.
Weaknesses
- Occlusions, unusual poses and crowded scenes still cause errors.
When to use itWhenever you need to understand body movement: sports, ergonomics, AR, sign language, animation.
Compare with
Check yourself
- Q01
What does a top-down pose estimator do first?
Show answer
A: Detect each person