IterativeClosestPoint (variants)

Introduction

Small experimental project for performing experiments with registration of surfaces using different approaches than the traditional Iterative Closest Point (ICP), which is used by vtkIterativeClosestPoint from the VTK repository. Many variants exists - as a start only only alternative is implemented. Instead of minimizing point distance, the point-to-plane distance is minimized, when registering the point $p$ to $q$.

$|(R(\theta)p + t - q)\cdot q|,$

where $R$ and $t$ is the rotation and translation, which are optimized.

This is a small snapshot of larger project based on the VTK repository and meant primarily for educational purposes.

Implementation

vtkImplicitPolyDataDistance2 A thread-safe version of the original vtkImplicitPolyDataDistance, where it is possible to return both the closest point and the gradient. It is pretty slow and should be optimized using meta-programming at some point.

vtkPolyDataCorrespondenceFilter A custom filter, which can be used for finding correspondences between two vtkPolyData. It is implemented using vtkSMPTools to speed-up the performance.

vtkICP Registration algorithm inheriting from vtkLinearTransform similarly to vtkIterativeClosestPoint. Right now, a point-to-point metric, a point-to-plane metric as well as a symmetrized version of the point-to-plane metric by Szymon Rusinkiewicz symmetrized can be chosen. The first algorithm is optimized using the standard closed form SVD and the latter ones are optimized using simple steepest descent. A Gauss-Newton version will follow at some point supporting various priors. For both algorithm a distance threshold can be set as well as the maximum number of landmarks.

Currently, we use the interpolated normal using vtkImplicitPolyDataDistance. We should return the interpolated normal instead.

Build and run

It is implemented following the directory and build structure of VTK. No attention has been given to performance (yet).

Example

There is a small Python example in RegistrationDemo demonstrating how registration using a point-to-plane metric converges in much fewer iterations than the conventional point-to-point metric.