Keypoint Detection

Find every person and draw their skeleton — two models running back to back on your own GPU, one to find people and one to pose them.

About this page

Human pose estimation locates the joints of each person in an image — shoulders, elbows, wrists, hips, knees and ankles — and joins them into a stick-figure skeleton. This page does keypoint detection for the 17 COCO keypoints on a photo, a sample or a live webcam feed, entirely in your browser.

It is top-down pose estimation, so two models run one after the other: a detector finds the people, then ViTPose base estimates the keypoints on each person's crop. The default pair is D-FINE nano with ViTPose base, a 180 MB combined download at fp16 on WebGPU or 92 MB quantized for the CPU, and it is fast enough for the webcam. The alternative pairs RT-DETR r50 with the same pose model, 260 MB, a stronger detector that runs on WebGPU only and is offered for still images. The picker quotes the combined size, because neither half works alone.

A person-confidence slider and a cap on how many people to pose travel with each run: every extra person is another pass through the pose model, so the cap keeps a crowded scene from stalling the tab. Each joint is drawn with an opacity that follows the model's confidence, because a heatmap always peaks somewhere and an occluded ankle comes back as a guess. Detection and pose timings are reported separately.

Both models run in a Web Worker on your device, and neither the photo nor the camera frames are uploaded.