HFS gives different results with CPU and GPU

[arrufat]

System information (version)

• OpenCV => 4.2.0
• Operating System / Platform => Arch Linux
• Compiler => GCC 9.3.0
• CUDA => 10.2.89

Detailed description

I know it has been discussed in #1417 and in opencv/opencv#15183.
However, I am experiencing the same problem on my machine.

Steps to reproduce

Compile on Arch Linux with the flags they used for OpenCV and enable CUDA:

build() {
  cd build
  export JAVA_HOME="/usr/lib/jvm/default"
  # cmake's FindLAPACK doesn't add cblas to LAPACK_LIBRARIES, so we need to specify them manually
  _pythonpath=`python -c "from sysconfig import get_path; print(get_path('platlib'))"`
  cmake -G Ninja ../$pkgname-$pkgver \
    -DWITH_OPENCL=ON \
    -DWITH_OPENGL=ON \
    -DWITH_TBB=ON \
    -DWITH_QT=ON \
    -DBUILD_WITH_DEBUG_INFO=OFF \
    -DBUILD_TESTS=OFF \
    -DBUILD_PERF_TESTS=OFF \
    -DBUILD_EXAMPLES=ON \
    -DINSTALL_C_EXAMPLES=ON \
    -DINSTALL_PYTHON_EXAMPLES=ON \
    -DCMAKE_INSTALL_PREFIX=/usr \
    -DCPU_BASELINE_DISABLE=SSE3 \
    -DCPU_BASELINE_REQUIRE=SSE2 \
    -DOPENCV_EXTRA_MODULES_PATH="$srcdir/opencv_contrib-$pkgver/modules" \
    -DOPENCV_SKIP_PYTHON_LOADER=ON \
    -DOPENCV_PYTHON3_INSTALL_PATH=$_pythonpath \
    -DLAPACK_LIBRARIES="/usr/lib/liblapack.so;/usr/lib/libblas.so;/usr/lib/libcblas.so" \
    -DLAPACK_CBLAS_H="/usr/include/cblas.h" \
    -DLAPACK_LAPACKE_H="/usr/include/lapacke.h" \
    -DOPENCV_GENERATE_PKGCONFIG=ON \
    -DOPENCV_ENABLE_NONFREE=ON \
    -DOPENCV_JNI_INSTALL_PATH=lib \
    -DOPENCV_GENERATE_SETUPVARS=OFF \
    -DEIGEN_INCLUDE_PATH=/usr/include/eigen3 \
    -DWITH_CUDA=ON \
    -DWITH_CUDNN=ON \
    -DOPENCV_DNN_CUDA=ON \
    -DCUDA_GENERATION=Auto
  ninja
}

I just perform the segmentation normally with default parameters.

auto engine = cv::hfs::HfsSegment::create(image.rows, image.cols);
auto hfs_cpu = engine->performSegmentCpu(image);
auto hfs_gpu = engine->performSegmentGpu(image);

Then I get the following results on the sample data 000.jpg:

• CPU
  
• GPU
  

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues,
  answers.opencv.org, Stack Overflow, etc and have not found solution
• I updated to latest OpenCV version and the issue is still there
• There is reproducer code and related data files: videos, images, onnx, etc

[llschloesser]

I have the same CPU-GPU result mismatch on my system and I would also appreciate this fix. This segmentation algorithm runs significantly faster than most alternatives in OpenCV.

Relatedly, the CUDA SLIC implementation in HFS would best be moved to the CUDA image processing module as a stand-alone algorithm.

[llschloesser]

It would also make sense for the HFS algorithm to accept segmentation labels as input rather than produce one internally through the hidden CUDA SLIC implementation (though perhaps the learned weights are only suitable for SLIC?).