fast-human-reg

A library for fast human registration. It achieves a speed of 0.2s/example with bs=1, and can be 10 times faster if you batchfy it!

The pipeline is based on a feedforward autoencoder (CorrAE) and fast SMPL fitter.

Dependency

The code is tested on torch=1.12.1+cu121, cuda12.1, debian11. We recommend using conda environment:

conda create -n fast-reg python=3.8
conda activate fast-reg 

Required packages can be installed by:

pip install -r pre-requirements.txt # Install pytorch and other dependencies 

SMPL body model, please download SMPL family models following this instruction.

Example Usage

python demo.py dataset.file=$PWD/data/demo-kinect-pc.ply dataset.smpl_root=<your_smpl_model_root>

Note that the model is trained with the ProciGen dataset, hence the human (head to foot vector) aligns with the +y axis, see the figure of the demo point cloud below. If your data has a different head orientation, please rotate them before doing registration.

Performance

Our fast registration pipeline achieves similar accuracy as NICP, but much faster, see results below (bs=1):

BEHAVE kinect point clouds:

Method	v2v(cm)	Runtime(s)
NICP	4.12	48.62
Ours	5.22	0.24

MGN human scans:

Method	v2v(cm)	Runtime(s)
NICP	5.07	46.46
Ours	6.25	0.24

Citation

If you use our code, please cite:

@inproceedings{xie2024InterTrack,
    title = {InterTrack: Tracking Human Object Interaction without Object Templates},
    author = {Xie, Xianghui and Lenssen, Jan Eric and Pons-Moll, Gerard},
    booktitle = {International Conference on 3D Vision (3DV)},
    month = {March},
    year = {2025},
}

@article{sarandi24nlf,
    title = {Neural Localizer Fields for Continuous 3D Human Pose and Shape Estimation},
    author = {Sárándi, István and Pons-Moll, Gerard},
    journal = {Advances in Neural Information Processing Systems (NeurIPS)},
    year = {2024},
}