Non-rigid Point Cloud Registration with Neural Deformation Pyramid [Paper]

Hierarchical non-rigid registration of multiple scans

Scale variant non-rigid registration with Sim(3) warp field

Requirements

The code tested on python=3.8.10, pytorch=1.9.0 with the following packages:

• pytorch3d, open3d, opencv-python, tqdm, mayavi, easydict

Obtain the 4DMatch benchmark

• Download the train/val/4DMatch-F/4DLoMatch-F split, (google drive, 14G). We filter point cloud pairs with near-rigid motions from the original 4DMatch benchmark. 4DMatch-F & 4DLoMatch-F denote the filtered benchmark.
• Extract it and create a soft link under this repository.

  ln -s /path/to/4Dmatch  ./data

Reproduce the result of NDP (no-learned)

• Run

  python eval_nolearned.py --config config/NDP.yaml  

  To visualize the registration result, add --visualize.

Reproduce the result of LNDP (supervised)

• First download pre-trained point cloud matching and outlier rejection models (google drive, 271M). Move the models to correspondence/pretrained
• Install KPConv

  cd correspondence/cpp_wrappers; sh compile_wrappers.sh; cd ../..

• Finally run

  python eval_supervised.py --config config/LNDP.yaml  

  To visualize the registration result, add --visualize.

Run shape transfer example

python shape_transfer.py -s sim3_demo/AlienSoldier.ply -t sim3_demo/Ortiz.ply

Our related projects

Lepard: rabbityl/lepard
DeformingThings4D: rabbityl/DeformingThings4D

Citation

If you use our code please cite:

@article{li2022DeformationPyramid, 
    title={Non-rigid Point Cloud Registration with Neural Deformation Pyramid.}, 
    author={Yang Li and Tatsuya Harada},
    journal={arXiv preprint arXiv:2205.12796},
    year={2022}
}