Home

Awesome

cpd

GitHub Workflow Status

Coherent Point Drift (CPD) is a point-set registration algorithm, originally developed by Andriy Myronenko et al. This is a C++ library that runs CPD.

CPD can be compared to Iterative Closest Point, another point-set registration algorithm that is widely used. While ICP minimizes point-to-point distances, CPD uses a Gaussian Mixture Model to minimize the error between a point and all other points. If you're thinking that this is very computationally intensive, you're right — both the CPD algorithm and the underlying error calculations take a lot of time, which is why we've created fgt to speed up those Gauss transforms. We hope this library provides a freer and more performant alternative to the original reference Matlab implementation.

This library supports three variants of CPD:

Andriy's reference implementation comes with one other type of registration, nonrigid_lowrank, which is not implemented in the latest version of this library (yet) (see History for information on how to find and use a previous version of this library that has nonrigid_lowrank).

This code lives on Github. It has some Doxygen documentation.

Usage

Basic, default usage can be accomplished via some namespace-level methods:

#include <cpd/rigid.hpp>

int main(int argc, char** argv) {
    cpd::Matrix fixed = load_points_from_somewhere();
    cpd::Matrix moving = load_points_from_somewhere();
    cpd::RigidResult result = cpd::rigid(fixed, moving);
    return 0;
}

Configuration is possible via Rigid, Nonrigid, and Affine:

#include <cpd/rigid.hpp>

int main(int argc, char** argv) {
    cpd::Matrix fixed = load_points_from_somewhere();
    cpd::Matrix moving = load_points_from_somewhere();
    cpd::Rigid rigid;
    rigid.correspondence(true).outliers(0.2);
    cpd::RigidResult result = rigid.run(fixed, moving);
    return 0;
}

If cpd is built with the jsoncpp component (see examples/ for a demonstration of the CMake configuration), the results of the cpd run can be converted to json:

#include <iostream>
#include <cpd/jsoncpp.hpp>
#include <cpd/rigid.hpp>

int main(int argc, char** argv) {
    cpd::Matrix fixed = load_points_from_somewhere();
    cpd::Matrix moving = load_points_from_somewhere();
    cpd::RigidResult result = cpd::rigid(fixed, moving);
    std::cout << cpd::to_json(result) << std::endl;
    return 0;
}

See the code and the documentation to discover all possible options, transformation methods, and probability calculation methods.

Examples

See examples/ in this code repository for some basic usage examples, including examples of how to set up a downstream CMake project that depends on cpd.

Installation

cpd depends on and CMake and Eigen at build time only — no runtime dependencies. For additional speed, it can also built with fgt. For json output of results, it can be built with jsoncpp.

On OSX

If you're on a Mac, use homebrew and my tap to install:

brew tap gadomski/gadomski
brew install cpd

From source

Use the usual CMake build incantation:

mkdir build
cd build
cmake ..
make

If you're using a home-built version of jsoncpp, make sure you set the following options when building and installing jsoncpp (this allows cpd to find jsoncpp):

Using downstream

cpd provides CMake export targets that you can import and use in your own project:

find_package(Cpd REQUIRED)
add_library(my-great-library
    the_code.cpp
    )
target_link_libraries(my-great-library
    PUBLIC
    Cpd::Library-C++
    )

The Cpd::Library-C++ target includes all the interface settings you need, so you shouldn't need any other calls to get set up.

If you'd like to enable json support, use the jsoncpp component:

find_package(Cpd COMPONENTS jsoncpp REQUIRED)
add_library(my-great-library the_code.cpp)
target_link_libraries(my-great-library PUBLIC Cpd::Library-C++ Cpd::Jsoncpp)

OpenMP

Both fgt and Eigen support OpenMP for some operations. As of yet, the interaction between the two is untested, so our official recommendation is to only use OpenMP with one of the projects, not both. If you do some work with OpenMP we'd love to hear how it goes.

Contributing

Github issues and pull requests, per usual.

History

The v0.1 and v0.2 lineages of cpd used armadillo for linear arithmetic instead of Eigen. Armadillo is a bit smoother for doing advanced eigenvalue decompositions and other operations, which made it a bit easier at first to directly port the Matlab reference implementation. For a couple of reasons, we decided to switch to Eigen for v0.3.

First, the Armadillo project had the bad habit of removing old versions from their download site, making it hard to maintain working code as their codebase developed. Second, many downstream applications use Eigen themselves, making Eigen a lower-friction choice for those users.

As of this writing, the Eigen implementation is less feature-full than the old Armadillo implementation, particularly with respect to the nonrigid_lowrank version. If you require some of that old functionality, use the v0.2 branch. If you need armadillo-5.x, which is required for the old cpd but is no longer available from the armadillo website, you can use my mirror. Thanks for your understanding during this switch.

Publications

This library has been used in the following publications:

License

This library is GPL2, copyright 2017 Peter J. Gadomski. See LICENSE.txt for the full license text.

This work is directly inspired by Andriy Myronenko's reference implementation, and we owe him many thanks.