Skip to content

GVHMR

Recover world-grounded human motion from monocular video

Visit official site
Category AI models

What GVHMR is

The model is described as a human motion recovery technique that uses a gravity-view coordinate system to recover world-grounded human motion from monocular video, with its training process and model weights made public.

What you can do with it

  • Use a gravity-view coordinate system to reduce ambiguity in image-to-pose mapping
  • Estimate pose positions in the gravity-view coordinate system to avoid accumulated error across consecutive frames
  • Convert estimated poses back to world coordinates to form a global motion sequence
  • Process per-frame features with a relative transformer and multi-task MLPs

Application scenarios

  • Recover precise 3D motion from actor performance video in action film production
  • Analyse athlete movement in sports training
  • Turn real-world motion data into game character animation in game development

Getting started

  • Visit the official site of the technique
  • Download and install the required software and dependency libraries
  • Prepare training datasets or test videos following the documentation
  • Use the provided code and weights to train on datasets or recover motion from video

Pros & cons

✓ What we like

  • Training is described on AMASS, BEDLAM, H36M and 3DPW datasets
  • Estimated poses can be converted back to world coordinates for a global motion sequence

! What to watch out for

  • The approach depends on a gravity-view coordinate system for pose estimation
  • Running the code requires installing software and dependencies locally

FAQ

What input does the technique work with?

The material describes recovery of world-grounded human motion from monocular video.

Which datasets are named for training?

AMASS, BEDLAM, H36M and 3DPW are listed as training datasets.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review