Skip to content

Latest commit

 

History

History
131 lines (113 loc) · 7.68 KB

MTF_setup_Ubuntu18_Matlab2018b.md

File metadata and controls

131 lines (113 loc) · 7.68 KB
Raw

Procedure for Installing Modular Tracking Framework (and OpenCV, etc.) on Ubuntu 18.04

Bernard Llanos

January 17, 2019

Remarks

  • The official installation instructions are at: https://github.com/abhineet123/MTF
  • I omit 'sudo' in the commands listed below; Add it when it makes sense.
  • My system has quite a few things installed for other programs, so it is possible that I have forgotten to list some of the dependencies for MTF.
  • Tokens to search and replace in this file:
    • FILEPATH: An appropriate location in your filesystem.

Prerequisites

  • Git: apt install git
  • CMake: apt install cmake
  • (For the Python 3 interface) Numpy: apt install python3-numpy
  • Eigen3: apt install libeigen3-dev
  • (For the MATLAB interface) MATLAB
    • I have R2018b.
  • Boost: apt install libboost-all-dev
  • HDF5: apt install libhdf5-openmpi-100 libhdf5-dev
    • I am not sure this is really a dependency, but the MTF build process did look for it. I might not have needed libhdf5-dev.
  • apt install liblz4-dev

OpenCV

  • For MTF to work with MATLAB, you should build MTF with the same version of OpenCV that is packaged with MATLAB. (Feel free to build a different copy of MTF to use outside of MATLAB with a different version of OpenCV.)
    • You can find out which version of OpenCV MATLAB has by running ls -l FILEPATH/R2018b/bin/glnxa64/*opencv*.
  • I haven't yet installed OpenCV's "contrib" modules; We'll see if they are necessary depending on which trackers you want to use from Modular Tracking Framework.

Dependencies

  • apt install libtbb-dev python3-dev
    • python3-dev may not be necessary; I probably already had the equivalent for Python 2.

Download

Building

  • Extract the source code archive.
  • cd FILEPATH/opencv-3.4.0
  • mkdir build
  • cd build
  • cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=FILEPATH ..
    • I had a compilation error regarding a missing header file that I solved by adding -DWITH_NVCUVID=OFF to the above command.
    • Here, you specify where the library should be installed.
    • (Optional) To use this version of OpenCV with Python, add what you set for CMAKE_INSTALL_PREFIX to your PYTHONPATH.
      • e.g. export PYTHONPATH=$PYTHONPATH:$CMAKE_INSTALL_PREFIX/python
  • make -j4
    • This builds both Python 2 and Python 3 libraries.
  • make install

FLANN

  • Download version 1.8.4 from http://www.cs.ubc.ca/research/flann/
  • Extract the archive, and open the resulting directory in a shell session.
  • (Probably optional) Let CMake know where to find MATLAB: export PATH=FILEPATH/R2018b/bin${PATH:+:${PATH}}
  • mkdir build, cd build, cmake ..
  • make
  • make install

VISP

  • This is an optional dependency.
  • The Ubuntu repositories now contain a version that is sufficiently up to date.
  • apt install libvisp-dev libvisp-doc visp-images-data
  • For some reason, I think I ended up not building MTF with VISP.

Caffe

  • Optional; I wouldn't install it unless you need the features that it provides. When I was building MTF, I think Caffe ended up being unused (even though I had told CMake where to find Caffe).
  • apt install caffe-cuda (if you have NVIDIA CUDA; For CPU version: apt install caffe-cpu)
  • Just in case, I installed some suggested packages: apt install libcaffe-cuda-dev python3-netcdf4 vitables

Modular Tracking Framework

Building

  • Clone the repository from: https://github.com/abhineet123/MTF
  • I created a custom installation directory: 'FILEPATH/installation'
    • Subdirectories: 'bin', 'include', 'lib', 'matlab', and 'python'.
  • I added this directory to my environment variables in '~/.bashrc': 
    export LIBRARY_PATH=$LIBRARY_PATH:/FILEPATH/installation/lib/
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/FILEPATH/installation/lib/
export C_INCLUDE_PATH=$C_INCLUDE_PATH:/FILEPATH/installation/include/
export CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/FILEPATH/installation/include/
export PYTHONPATH=$PYTHONPATH:/FILEPATH/installation/python
  • Open the repository directory in a shell session.
  • Edit 'CMakeLists.txt' so that the right version of OpenCV will be used: Replace "find_package(OpenCV REQUIRED)" with "find_package(OpenCV REQUIRED PATHS FILEPATH/opencv-3.4.0/install NO_DEFAULT_PATH)"
  • mkdir build && cd build
  • cmake .. -DMTF_INSTALL_DIR=FILEPATH/installation -DPY_VER=3.6 -DPYTHON_INCLUDE_DIR=/usr/include/python3.6 -DPYTHON_LIBRARY=/usr/lib/x86_64-linux-gnu/libpython3.6m.so -DMTF_PY_INSTALL_DIR=FILEPATH/installation/python -DMTF_MEX_INSTALL_DIR=FILEPATH/installation/matlab -DWITH_VISP=ON -DWITH_THIRD_PARTY=OFF
    • I set WITH_THIRD_PARTY=OFF because third party modules can have problems with OpenCV 3.x.
    • I set WITH_VISP=ON, but CMake said it went unused.
  • make -j4

Testing the C++ library

  • Run FILEPATH/installation/bin/runMTF

Testing the Python interface

  • Run python3 FILEPATH/Examples/python/runMTF2.py

Further steps for setting up the MATLAB interface

  • Before launching MATLAB, run export LD_LIBRARY_PATH=FILEPATH/opencv-3.4.0/install/lib:FILEPATH/installation/lib
    • This will make MATLAB link the MEX-file containing MTF's MATLAB interface with the version of OpenCV that MTF was built against, and with the MTF shared library.
  • Launch MATLAB (matlab &).
  • Add the folder containing the MEX-file ('mexMTF2.mexa64', for Linux systems) to your MATLAB path ('FILEPATH/installation/matlab').
  • Test if you can call MTF: >> success = mexMTF2('init','pipeline c img_source u');

If it didn't work

  • If the error is "undefined symbol", you haven't set LD_LIBRARY_PATH correctly.
  • If the error is "Gateway function is missing", the following steps solved the problem for me.

Building the MEX-file from within MATLAB

  • Install header files that, for some reason, weren't needed earlier: apt install libgflags-dev libgoogle-glog-dev libprotobuf-dev
  • Compile with most optional features disabled (do this inside the MATLAB shell): >> mex -c FILEPATH/Examples/cpp/mexMTF2.cc -IFILEPATH/installation/include -I/usr/include/eigen3 -IFILEPATH/opencv-3.4.0/install/include -DDISABLE_THIRD_PARTY_TRACKERS -DDISABLE_XVISION -DDISABLE_DFM -DDISABLE_VISP -DDISABLE_REGNET
    • The point is to specify include directories for the MTF, Eigen, and OpenCV header files.
    • This should create an object file, 'mexMTF2.o'.
  • In MATLAB, run, to create the MEX-file: >> mex mexMTF2.o -LFILEPATH/opencv-3.4.0/install/lib -LFILEPATH/installation/lib -Ll/usr/lib/x86_64-linux-gnu -L/usr/lib/x86_64-linux-gnu/hdf5/serial -lopencv_calib3d -lopencv_core -lopencv_cudaarithm -lopencv_cudabgsegm -lopencv_cudacodec -lopencv_cudafeatures2d -lopencv_cudafilters -lopencv_cudaimgproc -lopencv_cudalegacy -lopencv_cudaobjdetect -lopencv_cudaoptflow -lopencv_cudastereo -lopencv_cudawarping -lopencv_cudev -lopencv_dnn -lopencv_features2d -lopencv_flann -lopencv_highgui -lopencv_imgcodecs -lopencv_imgproc -lopencv_ml -lopencv_objdetect -lopencv_photo -lopencv_shape -lopencv_stitching -lopencv_superres -lopencv_video -lopencv_videoio -lopencv_videostab -lboost_random -lboost_filesystem -lboost_system -lboost_thread -lhdf5 -lpthread -lsz -lz -ldl -lmtf
    • You can tell what libraries need to be linked by running ldd mexMTF2.mexa64 (on the MEX-file which didn't work in MATLAB).
    • I have NVIDIA CUDA on my system. If you don't have it, you will probably not have to link the "*cuda*" libraries.
  • I probably could have repeated the above build steps with VISP and DFM enabled if I had specified the appropriate include and library directories, and library linking directives.