Pose estimation

Inferring the position and orientation of a body or object from sensor data, often from RGB or depth images.

Pose estimation recovers where joints and limbs are in space so software can animate avatars, count exercises, or plan robot grasps. Monocular RGB, multi-camera rigs, or depth cameras can supply the input.

Kinect (2010) estimated 20 joint positions by labeling each depth pixel with a body part using random forests trained on synthetic training data (article). Later deep learning models dominate many pose benchmarks, but Kinect proved the feature could ship at console latency.