Analysis of Iterative Closest Point Algorithms

Two non-overlapping dragon meshes aligned by ICP

Introduction

This is the project for the course 3D Scanning & Motion Capture at Technical University of Munich. The goal of this project is to compare different variants of ICP.

• Sampling: full, uniform sampling and normal space sampling.
• Correspondence methods: Normal Shooting, Normal Shooting with Colors, Nearest Neighbors and Nearest Neighbors with Colors.
• Objective functions: Point-to-Point, Point-to-Plane, Symmetric ICP (all three in both linearized and full forms).

Installation

Requirements

- flann
- freeimage
- Eigen
- Ceres
- yaml-cpp
* Open3D    # optional
* glog      # for windows
* lz4       # for macOS

Installing Open3D with something besides homebrew is a nightmare, so it is an optional part.
Cmake will automatically handle this and disable visualization in case Open3D was not found.

Building and running on macOS

Everything is available with homebrew.

brew install flann
brew install freeimage
brew install ceres-solver
brew install eigen
brew install yaml-cpp
brew install lz4

CMakeLists contains hints on installation paths and will notify if they need to be set explicitly.

mv CMakeLists_macos.txt CMakeLists.txt
mkdir build && cd build
cmake ..
make

Building and running on Windows and possibly Linux

glog is added to the requirements for Windows.

mv CMakeLists_win_linux.txt CMakeLists.txt
mkdir build && cd build
cmake ..
make

Docker

If you want to use Docker, use the following command to create a Docker image and install and compile the required dependencies.

docker build . -t 3dsmc-icp

Start the development environment

docker-compose up

Access the environment

http://localhost:8443/
Clone this repository into the Docker image.

Running

cd build
./icp ../configs/<path_to_config>.yaml

In the config you can set:

• sampling method
• correspondence method
• optimization objective
• whether to use color information
• other optimization settings
• data path
• output directory (experiment name)
• whether to visualize and which metrics to evaluate
  For a complete list of options plese refer to include/ICPConfiguration.h

# Task
runShapeICP: true
runSequenceICP: false
# ICP type
useLinearICP: true
# ICP objective(s)
useSymmetric: true
# Sampling
sampling: NSPACE
samplingRatio: 0.05
# Correspondence method (NN / SHOOT / NN_COLORS / SHOOT_COLORS)
correspondenceMethod: NN
# Other settings
matchingMaxDistance: 5
nbOfIterations: 10
# Error metrics
evaluateRMSENearest: true
evaluateTransforms: true
evaluateTime: true                      
# Experiment
experimentName: plane_linear_nn_partial  # Experiment name for folder name
dataDir: ../../Data/greyc_partial/       # Output directory

Results are automatically saved into results directory, separated by experiment name and different meshes.
You can combine and average metrics with tools in /visualize: there is a README.

After visualization you will get images similar to these.

Data

The datasets were used:

• GREYC 3D Colored Mesh Database (preprocessed)
• TUM RGB-D dataset
  Download the data and extract to the directory near the repository (this way example configs can be used without modifications).

Directory layout

icp
├── CMakeLists_macos.txt
├── CMakeLists_win_linux.txt
├── Dockerfile
├── README.md
├── src
│   ├── DataLoader.cpp                      # Dataloader for GREYC meshes
│   ├── Evaluator.cpp                       # Class that 
│   ├── ICPConfiguration.cpp                # Class for configuration
│   ├── ICPOptimizer.cpp
│   ├── ProcrustesAligner.cpp
│   ├── Search.cpp
│   ├── Utils.cpp
│   └── main.cpp
├── configs
│   └── ...                                 # Sample configs
├── docker
│   ├── Dockerfile
│   └── docker-compose.yaml
├── external                                # Data utils
│   └── data_utils
│       ├── Eigen.h
│       ├── FreeImageHelper.cpp
│       ├── FreeImageHelper.h
│       ├── PointCloud.h
│       ├── SimpleMesh.h
│       └── VirtualSensor.h
├── include                                 # Headers
│   ├── Constraints.h
│   ├── DataLoader.h
│   ├── Evaluator.h
│   ├── ICPConfiguration.h
│   ├── ICPOptimizer.h
│   ├── ProcrustesAligner.h
│   ├── Search.h
│   └── Utils.h
├── reproduce                                # Everything you need to reproduce results.
│   ├── conf
│   │   ├── ...
│   │   ...
│   └── ...
├── visualize                                 # Everything you need for visualization of results.
│   ├── README.md
│   ├── plot_shape_experiments.py
│   ├── plot_visualize.py
│   └── requirements.txt
└── assets                                    # Images from this README