alicevision

编译 §

环境：Ubuntu20.04

版本

1. 如果系统中预先有一些旧依赖。首先会碰到 cmake 的 include 目录正确，但是 lib 寻找目录错误的问题。详见这个 issue。根据这篇文章，通过设置 CMake 搜索的优先级，可以解决路径查询问题。解决方案。

2. 编译 aliceVision，为了省时，可以将项目分为两部分编译。

   1. 使用 CMakeLists 里面的内置方法对依赖进行编译。

   2. 在 CMakeLists 中关闭依赖编译后，再对项目进行单独编译。

3. Ubuntu20.04 默认版本为 3.16 尝试使用代理失败。但是换成官网源码编译的 3.22 后代理能正常工作了。

4. 不确定问题：将 CMAKE_PREFIX_PATH 设置成依赖安装目录后，CMake 会出现以下 warning:

   CMake Warning (dev) in src/CMakeLists.txt:
   Policy CMP0111 is not set: An imported target missing its location property
   fails during generation.  Run "cmake --help-policy CMP0111" for policy
   details.  Use the cmake_policy command to set the policy and suppress this
   warning.
   
   IMPORTED_LOCATION not set for imported target "assimp::assimp"
   configuration "Debug".
   

   进而导致 make 的规则中找不到 assimp 的依赖。

   Linux-x86_64/libaliceVision_mesh.so.2.4: assimp::assimp-NOTFOUND
   

   

   解决办法：按照 warning 提示添加 IMPORTED_LOCATION

   # /home/chunibyo/workspace/AliceVision/src/aliceVision/mesh/CMakeLists.txt
   set_target_properties(assimp::assimp PROPERTIES IMPORTED_LOCATION "${ASSIMP_LIBRARY_DIRS}/${ASSIMP_LIBRARIES}")

   

调试 §

aliceVision 分为前端和后端两个部分。

后端编译得到一系列可执行二进制文件，可以通过命令行调用，默认生成在 /home/chunibyo/workspace/AliceVision/build/Linux-x86_64 路径下。 前端通过 subprocess.Popen 对后端的二进制文件进行调用。

因为对 c++ 的调用方式子进程，所以很难直接 gdb 进 python 进程后对 c++ 进行 debug。解决办法是设法使后端的进程停住，然后在 CLion 中直接 attach 进去。

aliceVision_cameraInit §

处理照片的元信息。

命令：

aliceVision_cameraInit  --sensorDatabase "/home/chunibyo/workspace/AliceVision/src/aliceVision/sensorDB/cameraSensors.db" --defaultFieldOfView 45.0 --groupCameraFallback folder --allowedCameraModels pinhole,radial1,radial3,brown,fisheye4,fisheye1,3deanamorphic4,3deradial4,3declassicld --useInternalWhiteBalance True --viewIdMethod metadata --verboseLevel info --output "/home/chunibyo/workspace/meshroom_cache/tmp/MeshroomCache/CameraInit/db58d443c068acdfe5245d3670f4ba016ff52160/cameraInit.sfm" --allowSingleView 1 --input "/home/chunibyo/workspace/meshroom_cache/tmp/MeshroomCache/CameraInit/db58d443c068acdfe5245d3670f4ba016ff52160/viewpoints.sfm"

参数：

1. sensorDatabase：拍摄设备的先验硬件信息 txt 数据库。

2. groupCameraFallback：

   1. global

   2. folder

   3. image

rig：https://github.com/alicevision/meshroom/wiki/Multi-Camera-Rig

同时支持输入初始位姿。

aliceVision_featureExtraction §

主要计算代码

The SIFT detector is controlled mainly by two parameters: the peak threshold and the (non) edge threshold.

The peak threshold filters peaks of the DoG scale space that are too small (in absolute value). Peak_thresh 越大，特征点越少。

The edge threshold eliminates peaks of the DoG scale space whose curvature is too small (such peaks yield badly localized frames). edge_thresh 越大，特征越多。

1. 可选特征点

std::string EImageDescriberType_informations()
{
  return  "Describer types used to describe an image:\n"
          "* sift: Scale-invariant feature transform.\n"
          "* sift_float: SIFT stored as float.\n"
          "* sift_upright: SIFT with upright feature.\n"
          "* akaze: A-KAZE with floating point descriptors.\n"
          "* akaze_liop: A-KAZE with Local Intensity Order Pattern descriptors.\n"
          "* akaze_mldb: A-KAZE with Modified-Local Difference Binary descriptors.\n"
#if ALICEVISION_IS_DEFINED(ALICEVISION_HAVE_CCTAG)
          "* cctag3: Concentric circles markers with 3 crowns.\n"
          "* cctag4: Concentric circles markers with 4 crowns.\n"
#endif
#if ALICEVISION_IS_DEFINED(ALICEVISION_HAVE_APRILTAG)
          "* tag16h5: AprilTag family tag16h5.\n"
#endif
#if ALICEVISION_IS_DEFINED(ALICEVISION_HAVE_OPENCV)
#if ALICEVISION_IS_DEFINED(ALICEVISION_HAVE_OCVSIFT)
          "* sift_ocv: OpenCV implementation of SIFT describer.\n"
#endif
          "* akaze_ocv: OpenCV implementation of A-KAZE describer.\n"
#endif
          "";
}

2. ConfigurationPreset: 特征点预设，默认 NORMAL

/**
* @brief The preset to control the number of detected regions
*/
enum class EImageDescriberPreset
{
  LOW = 0
  , MEDIUM
  , NORMAL
  , HIGH
  , ULTRA
};

2. maxNbFeatures：手动设置特征点数目的方法

3. describerQuality: 手动设置特征点的质量。提高质量后会增加scale层数。It’s a trade-off between performance and precision.

4. contrastFiltering: 默认是 static，根据某种标准对关键点进行权重排序

    /**
    * @brief The method used to filter out features with too low constrast (that can be considered as noise).
    */
    enum class EFeatureConstrastFiltering
    {
        /// Use a fixed threshold for all the pixels
        Static = 0,
        /// Use a threshold for each image based on image statistics
        AdaptiveToMedianVariance,
        /// Disable contrast filtering
        NoFiltering,
        /// Grid sort by peak value per octave and by scale at the end
        GridSortOctaves,
        /// Grid sort by scale*peakValue per octave and at the end
        GridSort,
        /// Grid sort per scale steps and at the end (scale and then peak value)
        GridSortScaleSteps,
        /// Grid sort per octaves and at the end (scale and then peak value)
        GridSortOctaveSteps,
        /// Filter non-extrema peak values
        NonExtremaFiltering
    };

5. relativePeakThreshold: dynPeakTreshold = relativePeakThreshold * medianOfGradiants

6. gridFiltering: 指的是尽量不使得某个格子内的特征点过于密集。首先会根据 scale 或者 peak 等进行排序。分为留下的和舍弃两部分特征点。