Skip to content

Latest commit

 

History

History
92 lines (74 loc) · 5.41 KB

INSTALL.md

File metadata and controls

92 lines (74 loc) · 5.41 KB
Raw

Installation guide

In this file, we report the list of instructions for the installation of the online-detection-demo considering the MATLAB version of the detection algorithm. We start covering the list of the dependencies of the application and the instructions for the installation of the external repositories used. Finally, we provide instructions for compiling and installing the modules of the application.

Dependencies:

The list of dependencies is as follows:

  • YARP
  • OpenCV
  • Cuda
  • Python
  • Matlab
  • Matlab Parallel Computing Toolbox
  • Lua
  • Posix.signal
    • luarocks
    • luaposix
  • The following Python packages:
    • numpy (tested version: 1.13.1)
    • opencv-python (tested version: 3.3.0.10)
    • tensorflow-gpu (tested versions: 1.5.0 for Cuda 9.0 and 1.13.1 for Cuda 10.0)
  • Other required packages: OpenBLAS, Boost C++, Google Protobuf Buffers C++, Google Logging, Google Flags, LevelDB, HDF5, LMDB, Snappy

You can follow the installtion instructions in the official repositories of the first dependencies (namely, YARP, Opencv, Cuda, Matlab, Python and lua), while you can install the list of packages with the following commands:

sudo apt-get install libopenblas-dev libboost-all-dev libprotobuf-dev protobuf-compiler \
libgoogle-glog-dev libgflags-dev libleveldb-dev libhdf5-serial-dev liblmdb-dev libsnappy-dev
python3.5 -m pip install numpy==1.13.1
python3.5 -m pip install opencv-python==3.3.0.10
python3.5 -m pip install opencv-python==1.5.0

External repositories

The external repositories required for the matlab version of this application is as follows:

Faster R-CNN and Caffe installation

For this step, we prepared the script fetch_install_faster_rcnn.sh for your convinience. The script will clone the official repository of Faster R-CNN in the external folder, fetch the correct CAFFE version and compile both of them in the correct location.

Run the script as follows, substituing the variables $1 $2 $3 with, respectively, the matlab location, the cuda location and the number of jobs you want to use for compilation. This is an example: $1: /usr/local/MATLAB/R2019b $2: /usr/local/cuda $3: 16

./Scripts/fetch_install_faster_rcnn.sh $1 $2 $3

If using the script you get some errors or if you prefer to install Faster R-CNN by yourself, you will find the list of instructions at this link.

FALKON installation

For this step, we prepared the script fetch_install_falkon.sh for your convinience. The script will fetch the correct version of Falkon and compile it in the correct location. Run the script as follows:

./Scripts/fetch_install_falkon.sh

If using the script you get some errors or if you prefer to install Falkon by yourself, you will find the list of instructions at this link.

Re3 Tracker installation

For this step, we prepared the script fetch_install_re3.sh for your convinience. The script will fetch the correct version of the Re3 tracker and compile it in the correct location. Run the script as follows:

./Scripts/fetch_install_re3.sh

If using the script you get some errors or if you prefer to install the Re3 Tracker by yourself, you will find the list of instructions at this link.

Installation

Build matlab modules

Run the following scripts:

matlab -nodisplay -nosplash -nodesktop -r "run('modules/modules_matlab/online_detection_build.m');exit;"
matlab -nodisplay -nosplash -nodesktop -r "run('modules/modules_matlab/startup.m');exit;"

Build all other modules

Follow these instructions:

mkdir build
ccmake ../
make 
make install

Note: Please, consider that while doing Cmake you can flag different options, depending on the modules that you want to compile. The default options will allow you to compile the basic version of the demo. Please refer to this description for details about the different options

Setting up the system

The implemented detection algorithm allows to train a new model online, in just few seconds. It relies on Faster R-CNN for feature extraction and on FALKON + Minibootstrap procedure for classification (more details here).

You can use your own Faster R-CNN pretrained weights as feature extraction module but we made available the ones used for our experiments. You can download them running the following command in the folder of the repository.

./Scripts/fetch_model.sh