3D Human Pose Estimation

3D human pose estimation infers body-joint locations, articulated motion, or full-body meshes from 2D images and video, helping applications such as animation, robotics, and movement analysis. Common approaches lift 2D keypoints into 3D, use temporal Transformers or graph networks to exploit motion and skeletal structure, or combine calibrated multi-view geometry with learned models; newer systems also use diffusion priors and mesh foundation models. Research increasingly extends beyond RGB to LiDAR, radar, WiFi, inertial, pressure, and event sensors, often adding biomechanical constraints to improve global motion and physical plausibility. A central challenge is generalizing from controlled benchmarks to varied real-world settings with occlusion, unusual bodies and poses, and sparse or noisy observations, where accurate joint positions do not always ensure reliable depth, contact, or biomechanical measurements.