DiReg: Unsupervised Point Cloud Registration with Self-Distillation [Oral at BMVC24]

This is the companion code for the DiReg algorithm reported in the paper "Unsupervised Point Cloud Registration with Self-Distillation" by Christian Löwens et al. accepted as oral at BMVC 2024 [arxiv]. The code allows the users to reproduce and extend the results reported in the paper. Please cite the above work when reporting, reproducing or extending the results.

Purpose of the repository

This software is a research prototype, solely developed for and published as part of the publication of DiReg. It will neither be maintained nor monitored in any way.

Requirements

We tested our code with following environment:

• pip:
  • numpy=1.24.3
  • scipy=1.10.1
  • matplotlib=3.7.1
  • open3d=0.13.0
  • tensorboard=2.14.0
  • tensorboardX=2.6.2.2
  • future-fstrings=1.2.0
  • easydict=1.11
  • joblib=1.3.2
  • scikit-learn=1.3.0
  • configargparse=1.7
  • minkowskiengine=0.5.4 (see their installation instructions)
  • pytorch3d=0.7.2 (see their requirements and installation instructions)
  • pyyaml=6.0.1
• conda-forge:
  • cudatoolkit=11.3.1
• pytorch
  • pytorch=1.11.0

3DMatch Dataset

Since our code builds on the repository of Self-supervised Geometric Perception (SGP), please use their reorganized 3DMatch dataset for training and testing. For this, please follow their instructions in theNded/SGP/code/README.md and adjust the dataset_path in the corresponding config in code/perception3d (see next sections).

Initial correspondences $\cal{C}_{\mathrm{raw}}$	Refined correspondences $\cal{C}_{\mathrm{ref}}$

Training

DiReg

To train our model, execute:

python code/perception3d/train.py --config code/perception3d/config_direg.yml

Make sure all paths in the config are valid. For our proposed alternative without momentum teacher use config_direg_no_momentum_teacher.yml instead. The configurable parameters can be found in adaptor.py or train.py.

Baselines

To train a baseline, use one of the other configs in code/perception3d except for SGP. In this case, please use sgp.py and config_sgp.yml instead.

Evaluation

To test our model or a baseline, execute:

python code/perception3d/train.py --config code/perception3d/config_test.yml

License

DiReg is open-sourced under the AGPL-3.0 license. See the LICENSE file for details.

For a list of other open source components included in DiReg, see the file 3rd-party-licenses.txt. Our own contribution is summarized as a single commit in this repository.