NICP: Neural ICP Fields for 3D Human Registration at Scale

Riccardo Marin · Enric Corona . Gerard Pons-Moll

@article{marin2024nicp,
      title={NICP: Neural ICP for 3D Human Registration at Scale}, 
      author={Riccardo Marin and Enric Corona and Gerard Pons-Moll},
      inproceedings={European Conference on Computer Vision},
      year={2024},
      organization={Springer}
}

This repository contains the inference code for NSR registration pipeline. The code has been tested on Linux Ubuntu 20.04.6 LTS, using python 3.8.13, and a GPU GeForce RTX 3080 Ti.

Getting Started

Clone the repo

git clone https://github.com/riccardomarin/NICP.git NICP
cd NICP

Create the environment

conda create -n nsr python=3.8.13
conda activate nsr

Run the installation script (it also contains checkpoint download)

./install.sh

You need to download the smplh model and place it in the support_data folder. The correct file structure is:

support_data
  |__ body_models
         |__smplh
             |__neutral
                 |__model.npz

Set the home directory path in ./src/lvd_templ/paths.py

home_dir                          = #e.g., '/home/ubuntu/Documents/NICP/'

You are ready to start!

Inference

To use NSR and fit all the scans into the demo folder, you can run the following command:

PYTHONPATH=. python ./src/lvd_templ/evaluation/evaluation_benchmark.py

We also provide a streamlit demo to run NSR on single shapes using a GUI.

PYTHONPATH=. streamlit run ./src/lvd_templ/evaluation/stream_demo.py

Change Settings in `evaluation_benchmark.py`

Depending on your use case, specify different parameters for the NSR pipeline. This can be easily done my command line. For example, this command will run a unidirectional chamfer distance refinement (-1: input with oultiers; 1: Partial input):

PYTHONPATH=. python ./src/lvd_templ/evaluation/evaluation_benchmark.py core.cham_bidir=-1

For visualization and inspection purposes, by default the output is provided in a canonical frame. If you want that the output is aligned with the original input, you can specify:

PYTHONPATH=. python ./src/lvd_templ/evaluation/evaluation_benchmark.py core.scaleback=True

If you don't know if the shape Y-axis is aligned, we also implemented a "rotation guessing" heuristic: it tries several different rotations for the shape and applies the one with the best NICP loss score. For example, the SCAPE shape in demo_guess_rot folder is not aligned, but the alignment provided by the heuristic is good enough to let NSR succeed:

PYTHONPATH=. python ./src/lvd_templ/evaluation/evaluation_benchmark.py core.challenge='guess_rot' core.guess_rot=True

If you want to characterize a run (and avoid overwrite), you can specify a tag which will be added to the output filenames:

PYTHONPATH=. python ./src/lvd_templ/evaluation/evaluation_benchmark.py core.tag='greedy' core.lr_ss=0.01

Others parameters (like the number of NICP iterations or learning rate) can be tuned in conf_test/default.yaml.

Example Evaluation: FAUST

To obtain NSR results of Table 2, you can run:

# Registering all the shapes
PYTHONPATH=. python ./src/lvd_templ/evaluation/evaluation_benchmark.py core.challenge='FAUST_train_reg','FAUST_train_scans' core.checkpoint='1ljjfnbx' -m

# Obtaining the p2p matching for the scans
PYTHONPATH=. python ./src/lvd_templ/evaluation/get_match.py core.regist=out_ss_cham_0 core.subdiv=2 core.challenge.name='FAUST_train_scans' core.checkpoint='1ljjfnbx'
PYTHONPATH=. python ./src/lvd_templ/evaluation/get_match.py core.regist=out_ss_cham_0 core.subdiv=0 core.challenge.name='FAUST_train_reg' core.checkpoint='1ljjfnbx'

# Evaluating the p2p matching and get avarage error
PYTHONPATH=. python ./src/lvd_templ/evaluation/get_error.py core.evaluate="_1ljjfnbx_out_ss_cham_0_0_" core.challenge.name='FAUST_train_reg'
PYTHONPATH=. python ./src/lvd_templ/evaluation/get_error.py core.evaluate="_1ljjfnbx_out_ss_cham_0_2_" core.challenge.name='FAUST_scan_reg'

Libraries updates might have a minor impact on the numbers.

Acknowledgments

Thanks to Garvita Tiwari for the proofreading and feedback, Ilya Petrov for code refactoring, and the whole RVH team for the support. The project was made possible by funding from the Carl Zeiss Foundation. This work is supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 409792180 (Emmy Noether Programme, project: Real Virtual Humans), German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. This project has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 101109330. Gerard Pons-Moll is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645.