pose estimation
/ POHZ es-tih-MAY-shun /
Pose estimation figures out how a body is arranged. Rather than just boxing a person, it locates the key joints — wrists, elbows, shoulders, hips, knees — and connects them into a stick-figure skeleton. From a single photo or video frame it answers: where are this person's hands, are they sitting or standing, are their arms raised? The same idea extends to hands (where every finger joint is), faces (eyes, nose, mouth corners), and even rigid objects (which way is the box facing).
Mechanically, it is a kind of fine-grained detection: the model is trained to spot a fixed set of named landmarks and predict each one's coordinates, then draw the bones between them. This drives fitness apps that count your squats and check your form, motion capture for games and films, sign-language and gesture interfaces, sports analytics breaking down a pitcher's mechanics, and physical-therapy tools tracking a patient's range of motion. Modern versions run live on a phone camera.
Be candid about where it breaks. Occlusion is the great enemy — when one limb hides behind another, or a person is half out of frame, the model must guess the hidden joints and often guesses wrong. Crowds and tangled limbs confuse which joint belongs to which body. Unusual poses, baggy clothing, odd camera angles, and poor lighting all degrade it. And a 2D pose gives joint positions in the flat image only; recovering true 3D body posture from a single ordinary camera is genuinely ambiguous and far less reliable than it looks.
A yoga app watches you through the phone camera and overlays a skeleton: dots on your shoulders, elbows, wrists, hips, knees, ankles, joined by lines. It checks the angle at your knee to tell you your lunge is too shallow. Turn sideways so one leg hides the other, and the skeleton's far leg starts to wobble and guess.
A skeleton of named joints — useful for form-checking, but it falters the moment a limb is hidden.
Pose estimation tracks geometry, not intent: it can tell you a hand is raised but not whether the person is waving, hailing a cab, or asking a question. And single-camera 3D pose is inherently ambiguous — depth must be guessed, so treat 3D outputs as estimates, not measurements.