Haiping Wang¹, Yuan Liu^*,2, Zhen Dong^†,1, Wenping Wang³

¹Wuhan University    ²The University of Hong Kong    ³Texas A&M University   
^*The first two authors contribute equally.    ^†Corresponding authors.   

In this paper, we propose a novel local descriptor-based framework, called You Only Hypothesize Once (YOHO), for the registration of two unaligned point clouds. In contrast to most existing local descriptors which rely on a fragile local reference frame to gain rotation invariance, the proposed descriptor achieves the rotation invariance by recent technologies of group equivariant feature learning, which brings more robustness to point density and noise. Meanwhile, the descriptor in YOHO also has a rotation equivariant part, which enables us to estimate the registration from just one correspondence hypothesis. Such property reduces the searching space for feasible transformations, thus greatly improves both the accuracy and the efficiency of YOHO. Extensive experiments show that YOHO achieves superior performances with much fewer needed RANSAC iterations on four widely-used datasets, the 3DMatch/3DLoMatch datasets, the ETH dataset and the WHU-TLS dataset.

• Preprint paper.
• Project page.
• Video.

• 2023-12-23: Quite simple scripts for YOHO/FCGF feature extraction on a give point cloud. Check how to use it at here.
• 2023-02-28: A multiview registration mehtod SGHR utilizing YOHO is accepted by CVPR 2023! 🎉🎉
• 2023-02-05: YOHO has been extended to IEEE TPAMI 2023, a.k.a, RoReg! 🎉🎉
• 2022-06-30: YOHO is accepted by ACM MM 2022! 🎉🎉
• 2021-09-01: The Preprint Paper is accessible on arXiv.
• 2021-07-06: YOHO using FCGF backbone is released.

Here we offer the FCGF backbone YOHO. Thus FCGF requirements need to be met:

• Ubuntu 14.04 or higher
• CUDA 11.1 or higher
• Python v3.7 or higher
• Pytorch v1.6 or higher
• MinkowskiEngine v0.5 or higher

Specifically, The code has been tested with:

• Ubuntu 16.04, CUDA 11.1, python 3.7.10, Pytorch 1.7.1, GeForce RTX 2080Ti.

• First, create the conda environment:

  conda create -n fcgf_yoho python=3.7
  conda activate fcgf_yoho
  

• Second, intall Pytorch. We have checked version 1.7.1 and other versions can be referred to Official Set.

  conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
  

• Third, install MinkowskiEngine for FCGF feature extraction, here we offer two ways according to MinkowskiEngine by using the version we offered:

  cd MinkowskiEngine
  conda install openblas-devel -c anaconda
  export CUDA_HOME=/usr/local/cuda-11.1 #We have checked cuda-11.1.
  python setup.py install --blas_include_dirs=${CONDA_PREFIX}/include --blas=openblas
  cd ..
  

  Or following official command installation:

  pip install git+https://github.com/NVIDIA/MinkowskiEngine.git
  

• Fourth, install other packages, here we use 0.8.0.0 version Open3d for Ubuntu 16.04:

  pip install -r requirements.txt
  

The datasets and pretrained weights have been uploaded to Google Cloud:

• 3DMatch_train;
• 3DMatch/3DLomatch;
• ETH;
• WHU-TLS;
• Pretrained Weights. (Already added to the main branch.)

Also, all data above can be downloaded in BaiduDisk(Code:0di4).

Datasets above contain the point clouds (.ply) and keypoints (.txt, 5000 per point cloud) files. Please place the data to ./data/origin_data following the example data structure as:

data/
├── origin_data/
    ├── 3dmatch/
    	└── kitchen/
            ├── PointCloud/
            	├── cloud_bin_0.ply
            	├── gt.log
            	└── gt.info
            └── Keypoints/
            	└── cloud_bin_0Keypoints.txt
    ├── 3dmatch_train/
    ├── ETH/
    └── WHU-TLS/

Pretrained weights we offer include FCGF Backbone, Part I and Part II. Which have been added to the main branch and organized following the structure as:

model/
├── Backbone/
	└── best_bal_checkpoint.pth
├── PartI_train/
	└── model_best.pth
└── PartII_train/
	└── model_best.pth

To train YOHO, the group input of train set should be prepared using the FCGF model we offer, which is trained with rotation argument in [0,50] deg, by command:

python YOHO_trainset.py

Warning: the process above needs 300G storage space.

The training of YOHO is two-stage, you can run which with the commands sequentially:

python Train.py --Part PartI
python Train.py --Part PartII

With the Pretrained/self-trained models, you can try YOHO with:

python YOHO_testset.py --dataset demo
python Demo.py

To evalute YOHO on 3DMatch and 3DLoMatch:

• Prepare the testset:

  python YOHO_testset.py --dataset 3dmatch --voxel_size 0.025
  

• Evaluate the results:

  python Test.py --Part PartI  --max_iter 1000 --dataset 3dmatch    #YOHO-C on 3DMatch
  python Test.py --Part PartI  --max_iter 1000 --dataset 3dLomatch  #YOHO-C on 3DLoMatch
  python Test.py --Part PartII --max_iter 1000 --dataset 3dmatch    #YOHO-O on 3DMatch
  python Test.py --Part PartII --max_iter 1000 --dataset 3dLomatch  #YOHO-O on 3DLoMatch
  

  Where PartI is YOHO-C and PartII is YOHO-O, max_iter is the ransac times, PartI should be run first. All results will be stored in ./data/YOHO_FCGF.

The generalization results on the outdoor ETH dataset can be got as follows:

• Prepare the testset:

  python YOHO_testset.py --dataset ETH --voxel_size 0.15
  

  If out of memory, you can

  • Change the parameter batch_size in YOHO_testset.py-->batch_feature_extraction()-->loader from 4 to 1
  • Carry out the command scene by scene by controlling the scene processed now in utils/dataset.py-->get_dataset_name()-->if name==ETH

• Evaluate the results:

  python Test.py --Part PartI  --max_iter 1000 --dataset ETH --ransac_d 0.2 --tau_2 0.2 --tau_3 0.5 #YOHO-C on ETH
  python Test.py --Part PartII --max_iter 1000 --dataset ETH --ransac_d 0.2 --tau_2 0.2 --tau_3 0.5 #YOHO-O on ETH
  

  All the results will be placed to ./data/YOHO_FCGF.

The generalization results on the outdoor WHU-TLS dataset can be got as follows:

• Prepare the testset:

  python YOHO_testset.py --dataset WHU-TLS --voxel_size 0.8
  

• Evaluate the results:

  python Test.py --Part PartI  --max_iter 1000 --dataset WHU-TLS --ransac_d 1 --tau_2 0.5 --tau_3 1 #YOHO-C on WHU-TLS
  python Test.py --Part PartII --max_iter 1000 --dataset WHU-TLS --ransac_d 1 --tau_2 0.5 --tau_3 1 #YOHO-O on WHU-TLS
  

  All the results will be placed to ./data/YOHO_FCGF.

I provide two quite simple scripts in simple_yoho but I have not fully checked.

• fcgf_feat.py can be used for FCGF feature extraction, the output is a set of down sampled points and their FCGF features, but you need to download the orignal checkpoints and set the corresponding file path to pth in fcgf_feat.py.
• yoho_extract.py can be used for YOHO feature extraction, the output is 5000 randomly sampled keypoints and their corresponding yoho features.
• NOTE: a key parameter you should carefully set for both algos is voxel_size:
  • voxel_size = 0.025*(a rough scale of your pcs)/3m, the explanation is in the following context. (For indoor scenes, just setting it to 0.025 is always ok.)
  • voxel_size should be set to the same for the source and target pcs.

To test YOHO on other datasets, or to implement YOHO using other backbones according to your own needs, please refer to Here.

Please consider citing YOHO if this program benefits your project

@inproceedings{wang2022you,
  title={You only hypothesize once: Point cloud registration with rotation-equivariant descriptors},
  author={Wang, Haiping and Liu, Yuan and Dong, Zhen and Wang, Wenping},
  booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
  pages={1630--1641},
  year={2022}
}

Welcome to take a look at the homepage of our research group WHU-USI3DV !

We sincerely thank the excellent projects:

• EMVN for the group details;
• FCGF for the backbone;
• 3DMatch for the 3DMatch dataset;
• Predator for the 3DLoMatch dataset;
• ETH for the ETH dataset;
• WHU-TLS for the WHU-TLS dataset;
• PerfectMatch for organizing the 3DMatch and ETH dataset.