GVHMR anchors monocular motion capture to gravity

Many monocular motion-capture systems recover plausible 3D poses in camera coordinates. Problems emerge when those poses are placed in a shared world coordinate system: the body may tilt, its trajectory may drift, and its feet may slide or penetrate the ground. Those errors limit animation, sports analysis, robotics, and other applications that depend on gravity, contact, and absolute movement.

GVHMR, developed by ZJU3DV, recovers world-grounded human motion from a single RGB video. Gravity-View Human Motion Recovery estimates each frame in a coordinate system derived from world gravity and the camera’s viewing direction, reducing the accumulated error produced by frame-to-frame orientation prediction.

A stable frame for every video frame

Autoregressive approaches predict motion deltas between adjacent frames and integrate them over time. Small orientation and translation errors accumulate across long clips, producing tilted bodies, drifting trajectories, and unstable ground contact.

GVHMR defines a Gravity-View coordinate frame for each video frame. Gravity supplies a stable vertical axis, while the projected camera-view direction supplies a horizontal reference. The model predicts absolute body orientation in that frame, and estimated camera rotations map the results into a consistent global sequence.

This representation targets three common failure modes:

• Incorrect body tilt caused by an unstable estimate of the vertical axis.
• Long-horizon trajectory drift caused by accumulated frame-to-frame errors.