MVL Stereo Toolbox
~~~~~~~~~~~~~~~~~~
(C) 2013-2015 Rok Mandeljc <[email protected]>
Machine Vision Laboratory
University of Ljubljana, Faculty of Electrical Engineering


1. Introduction
~~~~~~~~~~~~~~~
Machine Vision Laboratory (MVL) Stereo Toolbox is a GUI-based testing
and tuning tool for various aspects of a stereo camera system. It was
written as a part of a project we were working on, where we needed
to test several stereo methods with various parameters to find an optimal
solution to our problem.

The toolbox is largely inspired by BM Stereo Tuner utility written by
Martin Peris (http://blog.martinperis.com/2011/08/opencv-stereo-matching.html),
but aims to solve more problems associated with configuration and tuning
of a stereo pipeline. As such, main highlights are:

- programmable model of stereo pipeline, which provides control over
  several steps: image pair acquisition, stereo calibration/rectification,
  disparity computation and reprojection
- Qt-based GUI for controlling each of steps in stereo pipeline
- several image pair sources, with option to implement additional ones
  via plug-ins
- several stereo methods, with option to implement additional ones via
  plug-ins
- programmable and GUI-based configuration for image pair sources and
  stereo methods
- stereo calibration wizard, calibration import and export, export of
  images at each processing step, stereo method parameter import and export
- cross-platform, written in C++ and using Qt and OpenCV


2. Installation
~~~~~~~~~~~~~~~

MVL Stereo Toolbox has following dependencies:
- C++ compiler and CMake 3.2 or newer
- Qt 5.5 or later
- OpenCV 3.0.0

The above will build stereo pipeline model and toolbox GUI. In addition,
stereo method plugins for all CPU-based methods provided in OpenCV are
built, and two image pair sources are built - OpenCV-camera-based one
file-loading-based one.

Additional plugins for image pair sources and stereo methods have following
dependencies:
- unicap development libraries and headers for unicap image pair source (linux-only)
- DC1394 development libraries and headers for dc1394 image pair source (linux-only)
- libelas for Efficient LArge Scale stereo method plugin (linux-only)
- OpenCV with CUDA stereo module for CUDA-based stereo method plugins
- CUDA (at least 5.0) in combination with OpenCV CUDA module for modified CUDA-based reprojection method

To build the toolbox, use cmake; in unpacked source directory, issue:

mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr/local
make
make install

The above sequence configures, builds and installs the toolbox into
/usr/local and sets the default plugin path to /usr/local/lib{64}/mvl-stereo-pipeline.

It is also possible to run the toolbox from build directory without installing
it. In this case, plugin directory must be overriden in order for plugins to
be loaded:

MVL_STEREO_TOOLBOX_PLUGIN_DIR=$PWD ./toolbox/MVLStereoToolbox


2.1 OpenCV with GPU/CUDA support:
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
OpenCV provides a module with GPU-accelerated functions using NVIDIA
CUDA architecture. However, this module is missing from some OpenCV
installations (most notably, on my Fedora development box).

The toolbox' cmake script will automatically detect presence or absence
of OpenCV CUDA module and accordingly enable/disable corresponding plugins.

In case you build CUDA-enabled OpenCV yourself, you can override the version
that toolbox is built against. To do so, use -DOpenCV_DIR switch when issuing
cmake, e.g.:

cmake -DOpenCV_DIR=/opt/opencv-3.0.0-gpu/share/OpenCV <the-rest-of-options>


2.2 Installing libelas for ELAS stereo method plugin (linux-only):
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Obtain ELAS source from author's web page (requires your e-mail address):
http://www.cvlibs.net/download.php?file=libelas.zip

Unfortunately, the sources do not build an installable shared library;
to fix this, grab the following patch:
http://mvg.fe.uni-lj.si/~rokm/libelas-install-lib.patch

In the example below, libelas is installed in /opt/libelas; to install
it to default system directory, adjust CMAKE_INSTALL_PREFIX accordingly.

unzip Downloads/libelas.zip -d libelas
cd libelas
patch -p1 -i ../Downloads/libelas-install-lib.patch
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/opt/libelas ..
make
make install

Since the above sequence installed libelas to non-standard path,
PKG_CONFIG_PATH must be altered when issuing toolbox' cmake command:

PKG_CONFIG_PATH=/opt/libelas/lib64/pkgconfig cmake <the-rest-of-options>