zgojcic / OverlapPredator

Official implementation of "PREDATOR: Registration of 3D Point Clouds with Low Overlap".

PREDATOR: Registration of 3D Point Clouds with Low Overlap

This repository provides code and data required to train and evaluate PREDATOR, a model for pairwise point-cloud registration with deep attention to the overlap region. It represents the official implementation of the paper:

PREDATOR: Registration of 3D Point Clouds with Low Overlap

*Shengyu Huang, *Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, Konrad Schindler
|ETH Zurich |
* Equal contribution

Contact

If you have any questions, please let us know:

• Shengyu Huang {[e-mail]}
• Zan Gojcic {[e-mail]}

News

• 2020-11-30: Code and paper release

Instructions

This code has been tested on

• Python 3.6.9/3.7.4, PyTorch 1.4.0/1.5.1, CUDA 10.1/11.0, gcc 6.3/7.5, TITAN Xp/GeForce RTX 2080 Ti/GeForce GTX 1080Ti
• Python 3.8.5, PyTorch 1.8.0.dev20201124+cu110, CUDA 11.1, gcc 9.3.0, GeForce RTX 3090

Note: We observed random data loader crashes due to memory issues on machines with less than 64GB CPU RAM. If you observe similar issues, please consider reducing the number of workers.

Requirements

To create a virtual environment and install the required dependences please run:

git clone https://github.com/ShengyuH/OverlapPredator.git
virtualenv --no-site-packages predator -p python3; source predator/bin/activate
cd OverlapPredator; pip install -r requirements.txt
cd cpp_wrappers; sh compile_wrappers.sh; cd ..

in your working folder.

Datasets and pretrained models

We provide preprocessed 3DMatch pairwise datasets (voxel-grid subsampled fragments together with their ground truth transformation matrices), and two pretrained models on 3DMatch dataset. The preprocessed data and models can be downloaded by running:

sh scripts/download_data_weight.sh

Predator is the model evaluated in the paper whereas bigPredator is a wider network that is trained on a single GeForce RTX 3090.

Model	first_feats_dim	gnn_feats_dim	# parameters
Predator	128	256	7.43M
bigPredator	256	512	29.67M

The results of both Predator and bigPredator, obtained using the evaluation protocol described in the paper, are available in the bottom table:

Note: The pretrained models and processed data of ModelNet will be released in the following weeks.

Train

After creating the virtual environment and downloading the datasets, Predator can be trained using:

python main.py --mode train --exp_dir predator_3dmatch --first_feats_dim 128 --gnn_feats_dim 256

and biGPREDATOR using:

python main.py --mode train --exp_dir bigpredator_3dmatch --first_feats_dim 256 --gnn_feats_dim 512

Evaluate

To evaluate PREDATOR, the first step is to extract features and overlap/matachability scores by running:

python main.py --mode test --exp_dir predator_3dmatch --pretrain weights/Predator.pth --first_feats_dim 128 --gnn_feats_dim 256 --test_info 3DLoMatch

the features will be saved to snapshot/{exp_dir}/{test_info}. The estimation of the transformation parameters using RANSAC can then be carried out using:

python scripts/evaluate_predator.py --source_path snapshot/predator_3dmatch/3DLoMatch --n_points 1000 --benchmark 3DLoMatch --exp_dir est_3dlomatch_1000

dependent on n_points used by RANSAC, this might take a few minutes. The final results are stored in est_{test_info}_{n_points}/result. To evaluate PREDATOR on 3DMatch benchmark, simply replace 3DLoMatch by 3DMatch.

Demo

We prepared a small demo, which demonstrates the whole Predator pipeline using two random fragments from the 3DMatch dataset. To carry out the demo, please run:

python scripts/demo.py

The demo script will visualize input point clouds, inferred overlap regions, and point cloud aligned with the estimated transformation parameters.

Citation

If you find this code useful for your work or use it in your project, please consider citing:

@article{huang2020predator,
  title={PREDATOR: Registration of 3D Point Clouds with Low Overlap},
  author={Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, Konrad Schindler},
  journal={arXiv:2011.13005 [cs.CV]},
  year={2020}
}

Acknowledgments

In this project we use (parts of) the official implementations of the followin works:

• FCGF
• D3Feat
• 3DSmoothNet
• MultiviewReg
• SuperGlue
• DGCNN
• RPMNet
• KPConv

We thank the respective authors for open sourcing their methods.