Human Mesh Recovery

Human mesh recovery studies how to infer a person’s 3D body shape and articulated pose, often as a parametric mesh such as SMPL or SMPL-X, from images, video, or sensors. Methods use image encoders and transformers, sometimes combined with temporal or multi-view cues, pose priors, test-time fitting, or generative models to resolve ambiguity from occlusion and limited viewpoints. Work also extends to radar, pressure, and acoustic sensing, and to recovering people together with scenes, interactions, and identities for applications such as avatars, motion analysis, and monitoring. A central challenge is balancing accurate, anatomically and physically plausible meshes with reliable generalization across people, clothing, cameras, and unusual poses while keeping inference fast and metrically grounded.