AI Pose Estimator

About AI Pose Estimator

AI Pose Estimator runs MoveNet (Lightning and Thunder) from TensorFlow.js entirely in your browser on the WebGL backend. It detects a single person and returns 17 COCO-format 2D keypoints — nose, eyes, ears, shoulders, elbows, wrists, hips, knees and ankles — each with a confidence score. The model (~5MB for Lightning, ~12MB for Thunder) downloads once and is cached. On top of the raw keypoints the tool computes major joint flexion angles (elbow, shoulder, hip, knee) and lets you export everything as JSON or CSV, with both downscaled-canvas and original-resolution coordinates. All processing happens locally - no images, video or pose data are ever uploaded to a server.

Does this pose estimator send my webcam or photo to a server?

No. The AI Pose Estimator runs the entire pose-detection pipeline in your browser using MoveNet via TensorFlow.js. Your webcam stream or uploaded photo is decoded into a canvas in memory, the MoveNet keypoint detector runs on your own GPU (WebGL backend) or CPU, and the resulting 17 keypoints are drawn back onto the canvas — without a single byte leaving your device. There is no upload, no telemetry and no cloud inference. This is essential for fitness apps that should not stream your training videos to a third party, for posture screening where privacy matters, and for low-latency live overlays.

What 17 keypoints does MoveNet detect?

MoveNet returns the 17 COCO-format keypoints: nose, left/right eye, left/right ear, left/right shoulder, left/right elbow, left/right wrist, left/right hip, left/right knee, and left/right ankle. Each keypoint carries an x/y pixel coordinate and a confidence score from 0 to 1. There are no hand, foot or facial-mesh points and no z/depth value — the output is single-person 2D only. The confidence slider lets you hide low-confidence joints from the skeleton and from the joint-angle calculation.

Lightning vs Thunder — what is the difference?

Both are single-person MoveNet detectors that output the same 17 keypoints. Lightning (~5MB) is optimised for speed and is the default — ideal for real-time webcam tracking and mobile devices, often 30-60 fps on a recent laptop GPU. Thunder (~12MB) is a larger, more accurate model that is better for analysing still images where precision matters, at the cost of speed. You can switch between them in Detection Settings; the detector reloads on the next detection.

How accurate is MoveNet and what are the caveats?

MoveNet is accurate and reliable for clear, well-lit, mostly front-facing single-person shots, which is why it is popular for rep counting, form checks and posture work. The caveats matter for professional use: it is SINGLE-PERSON — it locates one body per frame and will not return multiple people; it is 2D only — there is no true 3D depth, so the z-axis cannot be measured; and accuracy degrades with occlusion (limbs hidden behind the torso), motion blur, extreme camera angles, loose clothing and poor lighting. Side views are harder than front views because hip and shoulder points overlap in 2D. Always sanity-check low-confidence joints.

How does the joint-angle analysis work?

After detection the tool computes the major joint flexion angles entirely client-side from the MoveNet keypoints using 3-point vector math (atan2): left/right elbow (shoulder–elbow–wrist), shoulder (elbow–shoulder–hip), hip (shoulder–hip–knee) and knee (hip–knee–ankle). An angle is only reported when all three of its keypoints pass the confidence threshold, so unreliable joints are skipped rather than guessed. Because the angles are measured in image space, keep the camera roughly perpendicular to the plane of motion for the most meaningful numbers — ideal for tracking squat depth, elbow lockout or knee flexion over time.

Can I export keypoints and joint angles as JSON or CSV?

Yes. Once a pose is detected you can download the annotated image as PNG, copy the structured data to your clipboard, or download a JSON or CSV file. Every export contains the 17 keypoints with confidence scores, the computed joint angles in degrees, the model used, and the confidence threshold. Critically, coordinates are provided BOTH in downscaled-canvas space and in original-resolution space (the tool stores the scale factor applied when large images are resized for display), so professional motion or biomechanics workflows can map results straight back onto the source image.

Does it work on mobile, and how large is the model download?

Yes — it works in any modern mobile browser with WebGL support. The MoveNet model downloads once (about 5MB for Lightning or 12MB for Thunder) and is then cached by the browser, so repeat detections are instant and continue to work offline. On phones, Lightning is recommended for the smoothest webcam frame rate. If WebGL is unavailable the tool falls back to the slower CPU backend and shows a clear error if the model cannot load at all.