GiNGR: Generalized Iterative Non-Rigid Point Cloud and Surface Registration Using Gaussian Process Regression.
BibTex:
@article{madsen2022gingr,
title={GiNGR: Generalized Iterative Non-Rigid Point Cloud and Surface Registration Using Gaussian Process Regression},
author={Madsen, Dennis and Aellen, Jonathan and Morel-Forster, Andreas and Vetter, Thomas and L{\"u}thi, Marcel},
journal={arXiv preprint arXiv:2203.09986},
year={2022}
}
The GiNGR framework allows one to perform non-rigid registration with an iterative algorithm that makes use of Gaussian Process Regression in each iteration to find its next state.
Existing algorithms can be converted into the GiNGR framework, and compared based on 3 properties:
- Kernel function: how similar should the deformation of neighbouring points be - this is determined based on their correlation
- Correspondence estimation function: how to estimate corresponding points between the moving instance (reference) and the target.
- Observation uncertainty: what is the noise assumption of the correspondence estimations?
This framework contains a general library to input these 3 properties.
The core part of the GiNGR framework is found in gingr/api/GingrAlgorithm
with the update
function performing one iteration of GiNGR update.
Different pre-implemented configuration files can be found in gingr/api/registration/config
for CPD and ICP.
To use the GiNGR framework, you can make use of the maven snapshot by including the follow in your Scala 3 script:
//> using repository "https://oss.sonatype.org/content/repositories/snapshots"
//> using lib "ch.unibas.cs.gravis::gingr:0.1.0-SNAPSHOT"
If using SBT, then add the following to your build.sbt
:
resolvers +=
"Sonatype OSS Snapshots" at "https://oss.sonatype.org/content/repositories/snapshots"
libraryDependencies += "ch.unibas.cs.gravis" %% "gingr" % "0.1.0-SNAPSHOT"
Or you can install GiNGR to your local repo (.ivy2) by running sbt publishLocal
.
To run the examples, we make use of the VSCODE IDE. For installation help, please see https://scalismo.org/docs/Setup/vscode.
After installing VSCODE:
- Go to the examples folder:
cd examples
- Setup Code IDE with
scala-cli setup-ide .
- Open the IDE with
code .
- Now run the individual examples from the "run" menu.
To use GiNGR, one need to specify the deformation model to use in form of a GPMM model as well as the correspondence estimation function and the uncertainty update.
The creation of the GPMM is separate from the registration step. For examples, look in the examples
folder where demo scripts have been created to compute and visualize GPMMs for an Armadillo, Bunny and a Femur bone. In the UI, the deformation model can be evaluated by sampling from it.
The next step is to define the correspondence and uncertainty estimation update.
For this, default configurations have been implemented for CPD and ICP.
Simple Demo applications can be found in examples/DemoICP
and examples/DemoCPD
The demo scripts both perform deterministic and probabilistic registration one after the other.
In each iteration a new GiNGR state is computed which contains the GPMM model, the current fit
, the target as well as all the GPMM model parameters (non-rigid and global pose).
The probabilistic implementation is based on the ICP-Proposal repository: https://github.com/unibas-gravis/icp-proposal
The posterior output from the ICP probabilistic registration of the femur bone can be visualized with apps/registration/DemoPosteriorVisualizationFemur
.
In examples/DemoLandmarks
we compare 10 iterations of CPD for the Armadillo with and without the use of landmarks.
In examples/DemoMultiResolution
we perform 3 different registrations of GiNGR one after the other.
First CPD is used on a very coarse mesh (100 vertices), then CPD is used on a medium fine mesh (500 vertices) and finally, ICP is used on a finer mesh (1000 points) to get the fine details of the target mesh.
In the GiNGR code base, the basic implementations of existing algorithms can also be found for comparison. The algoriths are found under gingr/other/algorithms
Implementation of the CPD algorithm from https://arxiv.org/pdf/0905.2635.pdf
Implementation of the BCPD algorithm from https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8985307
Implementation of the non-rigid ICP algorithms from https://gravis.dmi.unibas.ch/publications/2007/CVPR07_Amberg.pdf