Author: Florian Lier [flier AT techfak.uni-bielefeld DOT de]

CLF Perception vision is a collection of nodes that provide information for, e.g., robots based on color and depth images.

For now, this is considered a sandbox.

Install the latest NVIDIA cuda toolkit: https://developer.nvidia.com/cuda-toolkit

• This version (master) has been tested with version 8.0
• Install OpenCV minimum 3.3.1 with CUDA support AND OpenCV contrib modules
• You will also need to clone opencv_contrib inside the opencv folder see: "-D OPENCV_EXTRA_MODULES_PATH=../opencv_contrib/modules" below.

Invoke OpenCV's cmake with:

    -D WITH_CUDA=ON
    -D BUILD_opencv_face=ON

Here's our exact CMAKE invocation

    cmake  \
        -D CMAKE_BUILD_TYPE=RELEASE \
        -D BUILD_TESTS=OFF \
        -D BUILD_PERF_TESTS=OFF \
        -D WITH_QT=ON \
        -D WITH_GSTREAMER=OFF \
        -D WITH_OPENGL=ON \
        -D WITH_OPENCL=OFF \
        -D WITH_V4L=ON \
        -D WITH_TBB=ON \
        -D WITH_XIMEA=OFF \
        -D WITH_IPP=ON \
        -D BUILD_opencv_adas=OFF \
        -D BUILD_opencv_bgsegm=OFF \
        -D BUILD_opencv_bioinspired=OFF \
        -D BUILD_opencv_ccalib=OFF \
        -D BUILD_opencv_datasets=OFF \
        -D BUILD_opencv_datasettools=OFF \
        -D BUILD_opencv_face=ON \
        -D BUILD_opencv_latentsvm=OFF \
        -D BUILD_opencv_line_descriptor=OFF \
        -D BUILD_opencv_matlab=OFF \
        -D BUILD_opencv_optflow=OFF \
        -D BUILD_opencv_reg=OFF \
        -D BUILD_opencv_saliency=OFF \
        -D BUILD_opencv_surface_matching=OFF \
        -D BUILD_opencv_text=OFF \
        -D BUILD_opencv_tracking=OFF \
        -D BUILD_opencv_xobjdetect=OFF \
        -D BUILD_opencv_xphoto=OFF \
        -D BUILD_opencv_stereo=OFF \
        -D BUILD_opencv_hdf=OFF \
        -D BUILD_opencv_cvv=OFF \
        -D BUILD_opencv_fuzzy=OFF \
        -D WITH_CUDA=ON \
        -D CUDA_GENERATION=Auto \
        -D ENABLE_FAST_MATH=1 \
        -D CUDA_FAST_MATH=1 \
        -D WITH_CUBLAS=1 \
        -D OPENCV_EXTRA_MODULES_PATH=../opencv_contrib/modules \
        -D CMAKE_INSTALL_PREFIX="/vol/pepper/systems/pepper-robocup-nightly" \
        -D CMAKE_BUILD_TYPE=RelWithDebInfo \
        -D CMAKE_CXX_FLAGS="${CXXFLAGS}" \
        -D CMAKE_C_FLAGS="${CFLAGS}" \
        -D CMAKE_SKIP_BUILD_RPATH=FALSE \
        -D CMAKE_BUILD_WITH_INSTALL_RPATH=FALSE \
        -D CMAKE_INSTALL_RPATH="/vol/pepper/systems/pepper-robocup-nightly/lib" \
        -D CMAKE_INSTALL_LIBDIR=lib \
     ..

Then:

    This needs to be redone after refactoring. SOON.

NOTE: Images larger than 1000px will be resized before displaying. The actual feature detection will be done on the original input size.

Examples can be found in the data folder.

    %YAML:1.0

    # Input topic
    input_ros_topic: "/pepper_robot/sink/front/image_raw"
    
    # Image locations
    targets:
     - "@PREFIX@/share/clf_perception_vision/data/example_images/milk.png"
     - "@PREFIX@/share/clf_perception_vision/data/example_images/coco.png"
     - "@PREFIX@/share/clf_perception_vision/data/example_images/pringels.png"
    
    # Corresponding labels
    labels:
     - "Milk"
     - "Coconut milk"
     - "Chips"
    
    # ORB ONLY, number of maximum keypoints applied to an image
    # Higher values may result in better detection but slow down
    # the process
    maxkeypoints_orb: 1500
    
    # Number of minimum matches between query image and current camera image
    # to accept a valid identification
    minmatches: 8
    
    # Number of maximum matches. This may affect fitting speed if set too high
    maxmatches: 1000
    
    # Distance of two consecutive keypoints in an image. Setting this too
    # high will result in inaccurate results
    detectionthreshold: 0.55
    
    # Algorithm to extract keypoints
    keypointalgo: SURF
    
    # Matching algorithm. Valid combinations are:
    # SURF+KNN, SURF+BF (evil slow), ORB+BF
    matcher: KNN
    
    # Draw window
    silent: false
    
    # Scale up camera image, makes things more robust but slower
    # Scale factor is 2. Using a higher resolution of your camera
    # is recommended.
    pyr_up: 1

TBD