A CUDA implementation of Segmentation SGM

Segmentation SGM is an implementation of segmentation algorithm that utilizes Semi-Global Matching.
From a given disparity map, it extracts segments that cover obstacles. It is implemented using C++ and CUDA.

• The detailed algorithm is here (currently Japanese only).

• CUDA
• OpenCV 3.0 or later
• CMake 3.10 or later

$ git clone https://github.com/fixstars/segmentation-sgm.git
$ cd segmentation-sgm
$ mkdir build
$ cd build
$ cmake ..
$ make

If you have installed libSGM, you can run sample/movie_with_libsgm and sample/benchmark by following command.

cmake -DWITH_LIBSGM=ON -DDCMAKE_MODULE_PATH=path/to/libSGM -DCMAKE_INSTALL_PREFIX=path/to/libSGM ..

./segmentation_sgm_sample_movie left-image-format right-image-format camera.xml

• left-image-format
  • the left image sequence
• right-image-format
  • the right image sequence
• camera.xml
  • the camera intrinsic and extrinsic parameters

./segmentation_sgm_sample_movie img_c0_%09d.pgm img_c1_%09d.pgm daimler_gt_stixel.xml

The sample images available at Daimler Urban Scene Segmentation Benchmark Dataset 2014 and Daimler Ground Truth Stixel Dataset are used to test the software.

The Segmentation SGM performance obtained from benchmark sample

The "SGM Team"
Fixstars Corporation

Apache License 2.0