← AI Terminology
Pose Estimation
Pose estimation is a computer vision task that detects and localises the key points (joints) of a human body or object in an image or video — producing a skeletal representation that captures position and orientation.
Human pose estimation outputs 2D or 3D coordinates of keypoints such as elbows, knees, and shoulders.
Human pose estimation outputs 2D or 3D coordinates of keypoints such as elbows, knees, and shoulders.
Why It Matters in AI
Pose estimation enables action recognition (is this person falling?), sports analytics (is the athlete's technique optimal?), sign language translation (what gesture is this?), virtual try-on (where is the person's body?), and motion capture for animation. It is also a key component of augmented reality (align virtual objects with human joints) and physical therapy AI (assess patient movement quality). Modern single-image 3D pose estimation makes motion capture accessible without expensive studio equipment.
Key Points
| Aspect | Description |
|---|---|
| 3D pose | (x, y, z) world coordinates — requires depth or multi-view cameras, or monocular lifting |
| Top-down | Detect persons first → crop → estimate pose per person — high accuracy, scales with person count |
| Bottom-up | Detect all keypoints globally → group into people — fast, handles crowds, fixed compute cost |
| Keypoints | COCO skeleton: 17 keypoints (nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles) |
| Key models | OpenPose, MediaPipe Pose, ViTPose, RTMPose — ordered from classic to state-of-art |
| 2D keypoints | (x, y) pixel coordinates of each joint in the image plane — standard for video applications |
Simple Analogy
A stick figure artist who, given a photo of a person, can draw the skeleton underneath — marking every joint's position. Pose estimation automates this: it draws the anatomically correct stick figure for every person in every frame of a video, at real-time speed.
Common Usage Examples
- MediaPipe Pose:
mp_pose.Pose().process(image)— 33 keypoints, runs on-device in real time from mmpose.apis import inference_topdown— MMPose top-down single-person pose- ViTPose:
ViTPoseConfig— ViT-based backbone achieving state-of-art on COCO pose benchmark - Sports analytics: Hawk-Eye, Second Spectrum — track player skeleton in broadcast footage
- Rehabilitation AI: Kaia Health app uses MediaPipe pose for physiotherapy exercise feedback
Summary
In short: Pose estimation detects body joint positions in images and video — the computer vision task enabling action recognition, sports analytics, motion capture, and physical therapy AI by providing a skeletal representation of human movement.