DO-CV / sara

Easy-to-Use C++ Computer Vision Library

Sara: C++ Computer Vision Library

I HAVE MOVED TO GITLAB, PLEASE GOTO: https://gitlab.com/DO-CV/sara

SO LONG GITHUB!

Sara (सार) is a Sanskrit word meaning essence.

Sara tries to focus on:

1. having an easy-to-use and simple API;
2. having easy-to-understand and efficient implementations of computer vision algorithms;

   • well "efficient" if you try to use my Halide or CUDA backends for example. "Easy-to-read" thanks to the exciting advances in compiler technologies like Halide, so there is no need to explicitly manipulate CPU intrinsics on the mainstream architectures. I have documented that some of my image processing implementations perform much better than OpenCV's CPU/GPU implementations, much to my surprise... See here.

   • Also Sara can also decode videos by using nVidia's hardware acceleration and it is indeed very fast on 4K videos (3840x2160).

3. rigorous sufficiently good testing
   • well... now as much as my limited time will permit it.

As a side note, as of 2021, I realize that yes... the ignorance of youth made me inflate my statements without realizing it. This project has always been a one-man only project. Now I have more experience and this has become a bit truer.

The design of Sara libraries is driven by the KISS principle. I enjoy maintaining Sara and making it evolve when I feel like it.

Sara is licensed with the Mozilla Public License version 2.0.

Since June 12, 2019, Sara is now a C++ 17 only project. The improvements in the language are compelling enough to make the library more readable.

Supported compilers:

• Visual Studio 2017 and above
• gcc 7 and above

Documentation

I don't have much time to maintain the documentation. I'd love your help.

What I can easily do is to keep up-to-date the reference API documentation hosted in codedocs.xyz.

I also write documentation at readthedocs.org where I provide more mathematical details about my implementations. Still available but a bit outdated.

In any case you are always better off consulting the examples folder and the test folder.

The codes are generally short and carefully so they should help you to get up to speed with the library usage.

Why yet another library?

Of course, you should not use my library and use OpenCV instead. But keep on reading what I have to say.

Obviously I like crafting software from the ground up and understanding Computer Vision algorithms by reimplementing them from A to Z. Besides, not everybody in the industry likes and uses OpenCV.

Back then, I used OpenCV C API for the first time during my research internship at Siemens in 2008 and was not aware of its C++ API at all. It was a very frustrating experience especially for the inexperienced programmer that I was back then.

After a while, I started writing the library as a hobby to have a more easy-to-use library and also to gain a better mastery of the C++ language. Then the library Sara became more structured in 2009, when I started my PhD at the IMAGINE lab in Ecole des Ponts, ParisTech.

Parts of the library may be reused for applications in the industry as I do myself. The library is initially not designed for real-time critical applications. Some algorithms can scale well, some do not. In most use cases, the library should be fine otherwise.

Anyways realtimeness is a very large topic and very platform-dependent, in which case OpenCV may not be able to rescue you, especially in non mainstream platforms or low-power devices... However I do want to point out that Sara can offer the following:

• On CUDA platforms, some of my CUDA implementations are a good start and even better than OpenCV ones.
• For low-power devices, the Halide CPU backend provides a good start and we just need to optimize the scheduling code for non desktop architectures.

Time has flown. Years after years I try to keep the library alive but things happen a lot slowly.

Build the libraries

The information below is a bit outdated but the detailed steps are still useful. Your best bet is to have a look at the CI scripts like .gitlab-ci.yml or .travis.yml

To build the libraries, run:

1. Install the following packages:

   • On Debian-based distributions:

     sudo apt-get install -qq \
       cmake \
       doxygen \
       libjpeg8-dev \
       libpng12-dev \
       libtiff5-dev \
       libavcodec-ffmpeg-dev \
       libavformat-ffmpeg-dev \
       libavutil-ffmpeg-dev \
       qtbase5-dev
     
     # To install Python bindings.
     sudo apt-get install -qq \
       boost-python-dev \
       python3-dev
     

   • On Red Hat-based distributions:

     sudo yum install -y
       cmake \
       doxygen \
       libboost-test-dev \
       libjpeg-devel \
       libpng-devel \
       libtiff-devel \
       ffmpeg \
       ffmpeg-devel \
       qt-devel
     
     # To install Python bindings.
     sudo apt-get install -qq \
       libboost-python-dev \
       libpython3-devel
     

2. Build the library:

   mkdir build
   cd build
   cmake .. \
     -DCMAKE_BUILD_TYPE=Release \
     -DSARA_BUILD_SHARED_LIBS=ON \
     -DSARA_BUILD_SAMPLES=ON \
     -DSARA_BUILD_TESTS=ON
   make  -j`nproc`  # to build with all your CPU cores.
   

3. Run the tests to make sure everything is alright.

   ctest --output-on-failure
   

4. Create DEB and RPM package.

   make package
   

5. Deploy by install the Debian package with Ubuntu Software Center, or type:

   # Debian-based distros:
   sudo dpkg -i libDO-Sara-shared-{version}.deb
   
   # Red Hat-based distros:
   sudo rpm -i libDO-Sara-shared-{version}.deb