AdamMF908 / probreg

Python package for point cloud registration using probabilistic model (Coherent Point Drift, GMMReg, SVR, GMMTree, FilterReg, Bayesian CPD)

Probreg is a library that implements point cloud registration algorithms with probablistic model.

The point set registration algorithms using stochastic model are more robust than ICP(Iterative Closest Point). This package implements several algorithms using stochastic models and provides a simple interface with Open3D.

Core features

• Open3D interface
• Rigid and non-rigid transformation

Algorithms

• Maximum likelihood when the target or source point cloud is observation data
  • Coherent Point Drift(2010)
  • FilterReg(CVPR2019)
• Variational Bayesian inference
  • Bayesian Coherent Point Drift(2020)
• Distance minimization of two probabilistic distributions
  • GMMReg(2011)
  • Support Vector Registration(2015)
• Hierarchical Stocastic model
  • GMMTree(ECCV2018)

Transformations

type	CPD	SVR, GMMReg	GMMTree	FilterReg	BCPD (experimental)
Rigid	Scale + 6D pose	6D pose	6D pose	6D pose (Point-to-point, Point-to-plane, FPFH-based)	-
NonRigid	Affine, MCT	TPS	-	Deformable Kinematic (experimental)	Combined model (Rigid + Scale + NonRigid-term)

CUDA support

You need to install cupy.

pip install cupy

• Rigid CPD
• Affine CPD

Installation

You can install probreg using pip.

pip install probreg

Or install probreg from source.

git clone https://github.com/neka-nat/probreg.git --recursive
cd probreg
pip install -e .

Getting Started

This is a sample code that reads a PCD file and calls CPD registration. You can easily execute registrations from Open3D point cloud object and draw the results.

import copy
import numpy as np
import open3d as o3
from probreg import cpd

# load source and target point cloud
source = o3.io.read_point_cloud('bunny.pcd')
target = copy.deepcopy(source)
# transform target point cloud
th = np.deg2rad(30.0)
target.transform(np.array([[np.cos(th), -np.sin(th), 0.0, 0.0],
                           [np.sin(th), np.cos(th), 0.0, 0.0],
                           [0.0, 0.0, 1.0, 0.0],
                           [0.0, 0.0, 0.0, 1.0]]))
source = source.voxel_down_sample(voxel_size=0.005)
target = target.voxel_down_sample(voxel_size=0.005)

# compute cpd registration
tf_param, _, _ = cpd.registration_cpd(source, target)
result = copy.deepcopy(source)
result.points = tf_param.transform(result.points)

# draw result
source.paint_uniform_color([1, 0, 0])
target.paint_uniform_color([0, 1, 0])
result.paint_uniform_color([0, 0, 1])
o3.visualization.draw_geometries([source, target, result])

Resources

• Documentation

Results

Compare algorithms

CPD	SVR	GMMTree	FilterReg

Noise test

ICP(Open3D)	CPD	FilterReg

Non rigid registration

CPD	SVR	Filterreg	BCPD

Feature based registration

FPFH FilterReg

Time measurement

Execute an example script for measuring time.

OMP_NUM_THREADS=1 python time_measurement.py

# Results [s]
# ICP(Open3D):  0.02030642901081592
# CPD:  3.6435861150093842
# SVR:  0.5795929960149806
# GMMTree:  0.34479290700983256
# FilterReg:  0.039795294986106455

Citing

@software{probreg,
    author = {{Kenta-Tanaka et al.}},
    title = {probreg},
    url = {https://probreg.readthedocs.io/en/latest/},
    version = {0.1.6},
    date = {2019-9-29},
}