This install guide explains how to install the structure from motion pipeline in Ubuntu 14.10.
- Setting up a virtual machine (optional)
- Installing packages from Ubuntu repositories
- Downloading and installing other tools
- Running an example
Creating a virtual machine is an optional step, for example, if you are using Windows. You can skip this step and go directly to Installing packages from Ubuntu repositories instead.
Download and install the latest VirtualBox if you have't already.
Download the latest Ubuntu iso.
Create an image in VirtualBox and install Ubuntu.
I configured virtualbox to use:
- 5000 MB memory
- virtual harddisk
- type: VDI
- dynamically allocated storage
- 64 GB diskspace
After the VM has been created, you can set the following properties:
- 2 cores
- 128 MB video memory
- 3D acceleration enabled
These options are mainly determined by the limitations of the machine you run on (this is about as much as my laptop can handle). Generally, using more cores and more memory is a good idea.
Start the virtual machine, install Ubuntu from the iso we just downloaded. Tick the box about downloading any available updates when installing.
We also installed the following optional packages:
sudo apt-get install virtualbox-guest-utils
sudo apt-get install virtualbox-guest-x11
Doing so allows you to share the clipboard between the host and the guest.
Once you have Ubuntu up and running we need to install the necessary tools and libraries. Open a terminal and install the following packages:
# git
# You will need git to clone the lastest versions of the structure from motion
# software from github:
sudo apt-get install git
# cmake
# You need cmake to generate the Makefiles needed to build the
# structure from motion software from github:
sudo apt-get install cmake
# gfortran
# You need a Fortran compiler to compile (parts of) the structure from motion software:
sudo apt-get install gfortran
# Glog
# a logging library from google (https://github.com/google/glog):
sudo apt-get install libgoogle-glog-dev
# Atlas
# The "Automatically Tuned Linear Algebra Software" provides C and Fortran77
# interfaces to a portably efficient BLAS implementation, as well as a few routines
# from LAPACK (http://math-atlas.sourceforge.net/):
sudo apt-get install libatlas-base-dev
# Eigen3
# C++ template library for linear algebra: matrices, vectors, numerical solvers,
# and related algorithms (http://eigen.tuxfamily.org).
sudo apt-get install libeigen3-dev
# SuiteSparse
# suite of sparse matrix algorithms (http://faculty.cse.tamu.edu/davis/suitesparse.html):
sudo apt-get install libsuitesparse-dev
# zlib
# a library implementing the deflate compression method found in gzip and PKZIP:
sudo apt-get install zlib1g-dev
# libjpeg
# a library implementing the loading of jpeg images:
sudo apt-get install libjpeg-dev
# libboost
# library with 'all the features you wanted in C++ but weren't there'
sudo apt-get install libboost-dev
# Python imaging library may not be installed by default on the
# lighter flavors of Ubuntu (e.g. Lubuntu 14.10 or Ubuntu 14.10 server)
sudo apt-get install python-pil
# for viewing the point clouds afterwards
sudo apt-get install meshlab
Our repository includes two other repositories as submodules:
To make sure you get the contents of the submodules when checking out the structure-from-motion repository, use the --recursive option to git clone:
cd ${HOME}
git clone --recursive https://github.com/NLeSC/structure-from-motion.git
The Ceres Solver is "an open source C++ library for modeling and solving large, complicated optimization problems. It is a feature rich, mature and performant library which has been used in production at Google since 2010." This solver is needed by the bundle adjustment step of the structure from motion (SfM) pipeline. We already installed the Ceres dependencies (originally described here) in the previous section, so now we can proceed to download and install Ceres:
cd ${HOME}/structure-from-motion
wget http://ceres-solver.org/ceres-solver-1.10.0.tar.gz
tar zxf ceres-solver-1.10.0.tar.gz
mkdir ceres-bin
cd ceres-bin
cmake ../ceres-solver-1.10.0
make -j3
make test
sudo make install
Bundler is a structure-from-motion (SfM) system for unordered image collections. Bundler takes a set of images, image features, and image matches as input, and produces a 3D reconstruction of camera and (sparse) scene geometry as output.
Next, compile bundler_sfm:
cd ${HOME}/structure-from-motion/bundler_sfm
make
PMVS2 is multi-view stereo software that takes a set of images and camera parameters (generated by bundler), and then reconstructs 3D structure of an object or a scene visible in the images. The software outputs a dense point cloud, that is, a set of oriented points where both the 3D coordinate and the surface normal are estimated for each point.
CMVS is software for clustering views for multi-view stereo. It is basically a pre-processor for PMVS2 that takes the output of bundler and generates one or more (optimized) configuration files for PMVS2. CMVS is normally used to split the PMVS2 processing in multiple independent parts, for example when creating a 3D reconstruction on the basis of thousands of images (which would be too much for PMVS2 to handle all at once). However, even when the number of images used is small, there is an advantage in using CMVS as it also removes unused images from the data set, and provides the order in which PMVS2 should process the images. This significantly reduces the processing time needed by PMVS2. The version we included in the structure-from-motion repository is a fork of pmoulon. It contains both CMVS and PVMS2, adds a cmake configuration, and contains several bug and performance fixes.
Compile CMVS/PMVS like this:
cd ${HOME}/structure-from-motion/cmvs-pmvs/program
mkdir build
cd build
cmake ..
make
The pipeline can be started by cd'ing into a data directory, and starting the 'run-sfm.py' script from there:
cd ${HOME}/structure-from-motion/examples/rock
python ../../run-sfm.py
On my laptop the example finished in 1 hour. After finishing, you can open a .ply file from ${HOME}/structure-from-motion/examples/rock/bundle with MeshLab to see the outcome.