tdillonmit / lepard Goto Github PK

4DMatch is a benchmark for matching and registration of partial point clouds with time-varying geometry. It is constructed using randomly selected 1761 sequences from DeformingThings4D. Below shows point cloud pairs with different overlap ratios.

We tested the code on python 3.8.10; Pytroch version '1.7.1' or '1.9.0+cu111'; GPU model GeForce RTX-2080 or Nvidia A100.

conda env create -f environment.yml
conda activate lepard
cd cpp_wrappers; sh compile_wrappers.sh; cd ..

• 4DMatch/4DLoMatch benchmark:

  • train/val/test split (15.6GB).
  • Raw depth images (4.5GB)

• 3DMatch/3DLoMatch benchmark (from the Predator paper): train/val/test split (4.8GB).

• Pretrained model: 4DMatch & 3DMatch (500M).

Download and extract the 4DMatch split to your custom folder. Then update the data_root in configs/train/4dmatch.yaml and configs/test/4dmatch.yaml

• Evaluate pre-trained

python main.py configs/test/4dmatch.yaml

(To switch between 4DMatch and 4DLoMatch benchmark, modify the split configuration in configs/test/4dmatch.yaml)

• Train from scratch

python main.py configs/train/4dmatch.yaml

An exmaple can be found here: Nonrigid-ICP-Pytorch

Download and extract the 3DMatch split to your custom folder. Then update the data_root in configs/train/3dmatch.yaml and configs/test/3dmatch.yaml

• Evaluate pre-trained

python main.py configs/test/3dmatch.yaml

(To switch between 3DMatch and 3DLoMatch benchmark, modify the split configuration in configs/test/3dmatch.yaml)

• Train from scratch

python main.py configs/train/3dmatch.yaml

If you use Lepard code or 4DMatch data please cite:

@article{lepard2021, 
    title={Lepard: Learning partial point cloud matching in rigid and deformable scenes.}, 
    author={Yang Li and Tatsuya Harada},
    journal={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2022}
}