FMPose3D: Monocular 3D Pose Estimation via Flow Matching
Official model weights for FMPose3D: monocular 3D pose estimation via flow matching by Ti Wang, Xiaohang Yu, and Mackenzie Weygandt Mathis.
| Paper | Project Page | GitHub | PyPI |
Overview
FMPose3D lifts 2D keypoints from a single image into 3D poses using flow matching — a generative technique based on ODE sampling. It generates multiple plausible 3D pose hypotheses in just a few steps, then aggregates them using a reprojection-based Bayesian module (RPEA) for accurate predictions, achieving state-of-the-art results on human and animal 3D pose benchmarks.
Available Checkpoints
| Filename | Skeleton | Joints | Training Data | Description |
|---|---|---|---|---|
fmpose3d_humans.pth | H36M | 17 | Human3.6M | Human 3D pose estimation |
fmpose3d_animals.pth | Animal3D | 26 | Animal3D | Quadruped animal 3D pose estimation |
Quick Start
pip install fmpose3d
Weights are downloaded automatically when using the Python API:
from fmpose3d import FMPose3DInference
# Human — weights auto-downloaded on first use
result = FMPose3DInference().predict("photo.jpg")
# Animal
result = FMPose3DInference.for_animals().predict("dog.jpg")
For the full API reference, see the GitHub repository.
Manual Download
from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="DeepLabCut/FMPose3D", filename="fmpose3d_humans.pth")
# or
path = hf_hub_download(repo_id="DeepLabCut/FMPose3D", filename="fmpose3d_animals.pth")
api = FMPose3DInference(model_weights_path=path)
Citation
@misc{wang2026fmpose3dmonocular3dpose,
title={FMPose3D: monocular 3D pose estimation via flow matching},
author={Ti Wang and Xiaohang Yu and Mackenzie Weygandt Mathis},
year={2026},
eprint={2602.05755},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.05755},
}
License
Model weights are for non-commerical use. Please contact the EPFL TTO for future information. We thank the Swiss National Science Foundation (SNSF Project # 320030-227871) and the Kavli Foundation for financial support.