This repository contains a Pytorch implementation of the point set registration method RLLReg. Details about the method can be found in the 3DV 2020 paper "Registration Loss Learning for Deep Probabilistic Point Set Registration".

ArXiv: [paper]

If you find the code useful, please cite using

@InProceedings{Lawin_2020_3DV,
    author = {Felix J\"aremo Lawin and Per-Erik Forss\'en},
    title = {Registration Loss Learning for Deep Probabilistic Point Set Registration},
    booktitle = {{IEEE/CVF} International Virtual Conference on 3D Vision ({3DV})},
    month = {November},
    year = {2020}} 

• Clone the repository: git clone https://github.com/felja633/RLLReg.git
• Create a conda environment and install the following dependencies:

conda create -n rllreg python=3.7
conda activate rllreg
conda install -y numpy pathlib mkl-include pyyaml
conda install -y pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch
conda install -y -c conda-forge cudatoolkit-dev
pip install easydict visdom
pip install git+https://github.com/jonbarron/robust_loss_pytorch
conda install -y -c open3d-admin open3d

• Install MinkowskiEngine with GPU support. Follow instructions for conda environment on https://github.com/NVIDIA/MinkowskiEngine

git clone https://github.com/NVIDIA/MinkowskiEngine.git
cd MinkowskiEngine
python setup.py install --cuda_home=/path/to/conda/rllreg 

• Install pytorch geometric from https://github.com/rusty1s/pytorch_geometric

pip install torch-scatter==latest+cu102 -f https://pytorch-geometric.com/whl/torch-1.6.0.html
pip install torch-sparse==latest+cu102 -f https://pytorch-geometric.com/whl/torch-1.6.0.html
pip install torch-cluster==latest+cu102 -f https://pytorch-geometric.com/whl/torch-1.6.0.html
pip install torch-spline-conv==latest+cu102 -f https://pytorch-geometric.com/whl/torch-1.6.0.html
pip install torch-geometric

Download and unpack Velodyne scans from http://www.cvlibs.net/download.php?file=data_odometry_velodyne.zip

Download RGB-D scenes from http://3dmatch.cs.princeton.edu/ using http://vision.princeton.edu/projects/2016/3DMatch/downloads/rgbd-datasets/download.sh and unpack the file. Download train.txt and test.txt. These contain the official train/test splits which can be found in the file https://vision.princeton.edu/projects/2016/3DMatch/downloads/rgbd-datasets/split.txt. Place these text files in the 3DMatch dataset folder.

Set up your local environment by setting the correct paths for your system in config.py. Here you should set the paths to the datasets and pre-trained models.

The following pre-trained models are available for download:

For the version trained with contrastive loss, use the following models from https://github.com/chrischoy/FCGF

To further enable comparisons to DGR, download the weights for 3DMatch and Kitti.

Place all pre-trained weights in the same folder and set pretrained_networks to the path of that folder in config.py.

Scripts for evaluation are available at experiments/. For an evaluation of pairwise registration as described in the paper run:

python experiments/evaluation_kitti.py

Scripts for training are available at experiments/. If you want to train RLLReg for pairwise registration run:

python experiments/train_rll_kitti.py

This repository also includes a pytorch version of Density Adaptive Point Set Registration (DARE) and Joint Registration of Multiple Point Clouds (JRMPC). Further, models/feature_reg_model_fcgf_fppsr.py and models/feature_reg_model_fpfh_fppsr.py contain pytorch implementations of FPPSR using FCGF and FPFH features respectively.

Under external/DeepGLobalRegistration the official implementation of DGR is located. The code is copied from the original repository but it is modified to use relative paths.

Felix Järemo Lawin

email: felix.lawin@gmail.com

• To compute the robust loss we use the code from https://github.com/jonbarron/robust_loss_pytorch

• We construct a backbone feature extractor based on the network from FCGF in https://github.com/chrischoy/FCGF

• In the comparison with DGR we use the code from https://github.com/chrischoy/DeepGlobalRegistration

• For the comparisons with ICP and FGR we use the implementations from Open3D http://www.open3d.org/

• The registration code was built upon the code for DARE in https://github.com/felja633/DARE

• The original implementation for JRMPC can be found in https://team.inria.fr/perception/research/jrmpc/

Special thanks go to Shivangi Srivastava who helped with initial implementations of the work!