You Only Hypothesize Once: Point Cloud Registration with Rotation-equivariant Descriptors

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.

News

2021.9.1 Paper is accessible on arXiv.paper
2021.8.29 The code of the PointNet backbone YOHO is released, which is poorer but highly generalizable. pn_yoho
2021.7.6 The code of the FCGF backbone YOHO is released. Project page

Performance

Network Structure

Requirements

Here we offer the FCGF backbone YOHO, so the 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

Installation

Create the anaconda environment:

conda create -n fcgf_yoho python=3.7
conda activate fcgf_yoho
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch 
#We have checked pytorch1.7.1 and you can get the pytorch from https://pytorch.org/get-started/previous-versions/ accordingly.

#Install MinkowskiEngine, here we offer two ways according to the https://github.com/NVIDIA/MinkowskiEngine.git
(1) pip install git+https://github.com/NVIDIA/MinkowskiEngine.git
(2) #Or use the version we offer.
    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 ..

pip install -r requirements.txt

KNN build:

cd knn_search/
export CUDA_HOME=/usr/local/cuda-11.1 #We have checked cuda-11.1.
python setup.py build_ext --inplace
cd ..

Data Preparation

We need the 3DMatch dataset (Train, Test) and the 3DLoMatch dataset (Test).

We offer the origin train dataset containing the point clouds (.ply) and keypoints (.txt, 5000 per point cloud) here TrainData. With which, you can train the YOHO yourself.

We offer the origin test datasets containing the point clouds (.ply) and keypoints (.txt, 5000 per point cloud) here 3dmatch/3dLomatch, ETH and WHU-TLS.

Please place the data to ./data/origin_data for organizing the data structure as:

data
origin_data
3dmatch
sun3d-home_at-home_at_scan1_2013_jan_1
Keypoints
PointCloud
3dmatch_train
bundlefusion-apt0
Keypoints
PointCloud
ETH
wood_autumn
Keypoints
PointCloud
WHU-TLS
Park
Keypoints
PointCloud

Train

To train YOHO yourself, you need to prepare the origin trainset with the backbone FCGF. We have retrained the FCGF with the rotation argument in [0,50] deg and the backbone model is in ./model/backbone. With the TrainData downloaded above, you can create the YOHO trainset with:

python YOHO_trainset.py

Warning: the process above needs 300G storage space.

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

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

We also offer the pretrained models in ./model/PartI_train and ./model/PartII_train. If the model above is demaged by accident(Runtime error: storage has wrong size), we offer another copy here.model

Demo

With the pretrained models, you can try YOHO by:

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

Test on the 3DMatch and 3DLoMatch

With the TestData downloaded above, the test on 3DMatch and 3DLoMatch can be done by:

Prepare the testset

python YOHO_testset.py --dataset 3dmatch

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

Generalize to the ETH dataset

With the TestData downloaded above, without any refinement of the model trained on the indoor 3DMatch dataset, the generalization result on the outdoor ETH dataset can be got by:

Prepare the testset [if out of memory, you can (1)change the parameter "batch_size" in YOHO_testset.py-->batch_feature_extraction()-->loader from 4 to 1 (2)or carry out the command scene by scene by controlling the scene processed now in utils/dataset.py-->get_dataset_name()-->if name==ETH]

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

Eval 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.

Generalize to the WHU-TLS dataset

With the TestData downloaded above, without any refinement of the model trained on the indoor 3DMatch dataset, the generalization result on the outdoor TLS dataset WHU-TLS can be got by:

Prepare the testset

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

Eval 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.