Pure Numpy Implementation of the Coherent Point Drift Algorithm


Keywords
image, processing, point, cloud, registration, mesh, surface, expectation-maximization, point-cloud, python
License
MIT
Install
pip install pycpd==1.0.3

Documentation

Python-CPD

https://travis-ci.com/siavashk/pycpd.svg?branch=master

Pure Numpy Implementation of the Coherent Point Drift Algorithm.

MIT License.

Introduction

This is a pure numpy implementation of the coherent point drift CPD algorithm by Myronenko and Song. It provides three registration methods for point clouds: 1) Scale and rigid registration; 2) Affine registration; and 3) Gaussian regularized non-rigid registration.

The CPD algorithm is a registration method for aligning two point clouds. In this method, the moving point cloud is modelled as a Gaussian Mixture Model (GMM) and the fixed point cloud are treated as observations from the GMM. The optimal transformation parameters maximze the Maximum A Posteriori (MAP) estimation that the observed point cloud is drawn from the GMM.

The registration methods work for 2D and 3D point clouds. For more information, please refer to my blog.

Pip Install

pip install pycpd

Installation From Source

Clone the repository to a location, referred to as the root folder. For example:

git clone https://github.com/siavashk/pycpd.git $HOME/pycpd

Install the package:

pip install .

For running sample registration examples under examples, you will need matplotlib to visualize the registration. This can be downloaded by running:

pip install matplotlib

Usage

Each registration method is contained within a single class inside the pycpd subfolder. To try out the registration, you can simply run:

python examples/fish_{Transform}_{Dimension}.py

where Transform is either rigid, affine or deformable and Dimension is either 2D or 3D. Note that examples are meant to be run from the root folder.