rancho86 / overlappredator

[CVPR 2021, Oral] PREDATOR: Registration of 3D Point Clouds with Low Overlap.

PREDATOR: Registration of 3D Point Clouds with Low Overlap (CVPR 2021, Oral)

This repository 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

For implementation using MinkowskiEngine backbone, please check this

For more information, please see the project website

Contact

If you have any questions, please let us know:

• Shengyu Huang {[email protected]}
• Zan Gojcic {[email protected]}

News

• 2021-03-25: Camera ready is on arXiv! I also gave a talk on Predator(中文), you can find the recording here: Bilibili, Youtube
• 2021-02-28: MinkowskiEngine-based PREDATOR release
• 2021-02-23: Modelnet and KITTI release
• 2020-11-30: Code and paper release

Instructions

This code has been tested on

• Python 3.8.5, PyTorch 1.7.1, CUDA 11.2, gcc 9.3.0, GeForce RTX 3090/GeForce GTX 1080Ti

Note: We observe random data loader crashes due to memory issues, if you observe similar issues, please consider reducing the number of workers or increasing CPU RAM. We now released a sparse convolution-based Predator, have a look here!

Requirements

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

git clone https://github.com/ShengyuH/OverlapPredator.git
virtualenv predator; 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

For KITTI dataset, please follow the instruction on KITTI Odometry website to download the KITTI odometry training set.

We provide

• preprocessed 3DMatch pairwise datasets (voxel-grid subsampled fragments together with their ground truth transformation matrices)
• raw dense 3DMatch datasets
• modelnet dataset
• pretrained models on 3DMatch, KITTI and Modelnet

The preprocessed data and models can be downloaded by running:

sh scripts/download_data_weight.sh

To download raw dense 3DMatch data, please run:

wget --no-check-certificate --show-progress https://share.phys.ethz.ch/~gsg/pairwise_reg/3dmatch.zip
unzip 3dmatch.zip

The folder is organised as follows:

• 3dmatch
  • train
    • 7-scenes-chess
      • fragments
        • cloud_bin_*.ply
        • ...
      • poses
        • cloud_bin_*.txt
        • ...
    • ...
  • test

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:

3DMatch(Indoor)

Train

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

python main.py configs/train/indoor.yaml

Evaluate

For 3DMatch, to reproduce Table 2 in our main paper, we first extract features and overlap/matachability scores by running:

python main.py configs/test/indoor.yaml

the features will be saved to snapshot/indoor/3DMatch. The estimation of the transformation parameters using RANSAC can then be carried out using:

for N_POINTS in 250 500 1000 2500 5000
do
  python scripts/evaluate_predator.py --source_path snapshot/indoor/3DMatch --n_points $N_POINTS --benchmark 3DMatch
done

dependent on n_points used by RANSAC, this might take a few minutes. The final results are stored in est_traj/{benchmark}/{n_points}/result. To evaluate PREDATOR on 3DLoMatch benchmark, please also change 3DMatch to 3DLoMatch in configs/test/indoor.yaml.

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 configs/test/indoor.yaml

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

KITTI(Outdoor)

We provide a small script to evaluate Predator on KITTI test set, after configuring KITTI dataset, please run:

python main.py configs/test/kitti.yaml

the results will be saved to the log file.

ModelNet(Synthetic)

We provide a small script to evaluate Predator on ModelNet test set, please run:

python main.py configs/test/modelnet.yaml

The rotation and translation errors could be better/worse than the reported ones due to randomness in RANSAC.

Custom dataset

We have a few tips for train/test on custom dataset

• If it's similar indoor scenes, please run demo.py first to check the generalisation ability before retraining
• Remember to voxel-downsample the data in your data loader, see kitti.py for reference

Citation

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

@inproceedings{predator,
  title={PREDATOR: Registration of 3D Point Clouds with Low Overlap},
  author={Shengyu Huang and Zan Gojcic and Mikhail Usvyatsov and Andreas Wieser, Konrad Schindler},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition, CVPR},
  year={2021}
}

Acknowledgments

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

• FCGF (KITTI preprocessing)
• D3Feat (KPConv backbone)
• 3DSmoothNet (3DMatch preparation)
• MultiviewReg (3DMatch benchmark)
• SuperGlue (Transformer part)
• DGCNN (self-gnn)
• RPMNet (ModelNet preprocessing and evaluation)

We thank the respective authors for open sourcing their methods. We would also like to thank reviewers, especially reviewer 2 for his/her valuable inputs.