libSGM_ocl V2.0

SGM implementation in OpenCL

Computes dense stereo correspondences.

Max disparities could be one of 64, 128, 256.

Using lower max disparity number increases performance and decreases memory usage.

4 path optimization optimizes disparity cost from left - > right, right -> left, up -> down, down-> up

8 path optimization adds oblique directions: upleft -> downright, upright -> downleft, downright -> upleft, downleft -> upright

If subpixel disparity is enabled, the result is multiplied by 16. You can calculate the floating point disparities by dividing the result by 16: float d = res / 16.0f;

Dependencies

• OpenCL

Install dependencies for ubuntu: sudo apt install ocl-icd-opencl-dev

On windows you can install OpenCL via vcpkg: vcpkg install opencl:x64-windows

Additional dependencies to build examples

• OpenCV

Ubuntu: sudo apt install libopencv-dev

Vcpkg: vcpkg install opencv:x64-windows

Building:

CMake options:

• BUILD_EXAMPLES - build examples, default value is ON
• CL_TARGET_OPENCL_VERSION - defines OpenCL target version, default value us 120

cmake .. -DBUILD_EXAMPLES=ON -DCL_TARGET_OPENCL_VERSION=120 -DCMAKE_BUILD_TYPE=Release

Note: if you are using vcpkg, please set either CMAKE_PREFIX_PATH or CMAKE_TOOLCHAIN_FILE as noted in vcpkg documentation.

cmake .. -DBUILD_EXAMPLES=ON -DCL_TARGET_OPENCL_VERSION=120 -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=[path to vcpkg]/scripts/buildsystems/vcpkg.cmake

Running stereo_movie example

Usage: stereo_movie.exe [params] img_source_left img_source_right

    --device_idx (value:0)
            OpenCL device index
    -h, --help (value:true)
            Print this message
    --max_disparity, --md (value:128)
            Maximum disparity
    --np, --num_path (value:4)
            Num path to optimize, 4 or 8
    --platform_idx (value:0)
            OpenCL plarform index
    --sp, --subpixel (value:false)
            Compute subpixel accuracy

    img_source_left (value:)
            Left images
    img_source_right (value:)
            Right images

Run on kitti rectified stereo database:

On Windows:

./stereo_movie [path to kitti]\sequences\00\image_0\000000.png [path to kitti]\00\image_1\000000.png 

On Linux:

./stereo_movie [path to kitti]/sequences/00/image_0/%06d.png [path to kitti]/00/image_1/%06d.png 

Known issues

• oblique path optimizations are not working correctly, there is some magic error. 4 path optimization gives better result now, 2 x faster and uses ~ half memory of 8 path optimization

Testing and performance

Kitti dataset, resolution 1241x376. Parameters:

• --subpixel=true
• --max_disparity=128
• --num_path=4

Nvidia Driver version: 455.32.00

AMD gpu driver, revision 20.40

Device	Windows 10 20H2	Ubuntu 2004
AMD Radeon RX 480 8GB	12 milliseconds	12 milliseconds
NVidia GTX 1070 8GB	8 milliseconds	7 milliseconds

Refs

Implementation based on libSGM:

• https://github.com/fixstars/libSGM

vcpkg:

• https://github.com/microsoft/vcpkg

Uses a modified version of CMakeRC:

• https://github.com/vector-of-bool/cmrc

Kitti:

• http://www.cvlibs.net/datasets/kitti/