Pose Estimation
Detecting the positions of joints (elbows, knees, etc.) or landmarks (eyes, nose) to infer the posture of a person or object.
Think of it like
Like marking the joints on a stick figure: head, shoulders, elbows, hips, knees. The model predicts where each joint is in the image.
Example
A fitness app analyzes a video of a user doing squats, detecting their knee, hip, and ankle positions to check form. A game uses hand pose to track gesture controls.
How it actually works
Human pose estimation typically predicts 2D pixel coordinates (and confidence) for each joint. Some models also infer 3D coordinates or depth. Challenging because people occlude themselves, appear at different scales, and in different lighting. Modern approaches use heatmaps (saliency maps at each joint) or direct regression. Top-down (detect person first, then their joints) vs. bottom-up (detect all joints, then link them).
For product teams
Powers fitness tracking, gesture control, animation from video, sports analytics, and accessibility interfaces.
For engineers
Standard datasets: COCO (17 joints), OpenPose (25 joints). Metrics: OKS (object keypoint similarity). Models: OpenPose, HRNet, MediaPipe.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome