Seeing Machines HMR: Real-time 3D Human Mesh Recovery at up to 180 FPS on NVIDIA Jetson Thor

Abstract

Robots working alongside people need timely information about human posture, movement and occupied space to make appropriate decisions as interactions unfold. Human Mesh Recovery (HMR), one of the leading features of Seeing Machines' Human-Centred Physical AI Platform, reconstructs 3D persons from a single-view RGB or RGB-D input in real time on embedded systems. Given a bounding box around a person, the system predicts articulated dense 3D pose, 3D body shape, and metric camera-frame placement, achieving state-of-the-art 3D pose estimation accuracy with significantly lower latency than frontier models such as Meta’s SAM-3D-Body and Fast-SAM-3D-Body. On NVIDIA Jetson Thor, HMR achieves 31.8 mm MPJPE on Harmony4D at 15.8 ms median latency with only 386M parameters: similar in accuracy to SAM-3D-Body with ~33× speed-up and ~70% fewer parameters. HMR-Lite achieves 5.5 ms median latency with only 146M parameters: similar in accuracy to Fast-SAM-3D-Body with ~20× speed-up and ~88% fewer parameters.