【复现笔记】Iterative Corresponding Geometry

文献：

M. Stoiber, M. Sundermeyer and R. Triebel, "Iterative Corresponding Geometry: Fusing Region and Depth for Highly Efficient 3D Tracking of Textureless Objects," 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 6845-6855, doi: 10.1109/CVPR52688.2022.00673.

论文网址：

Iterative Corresponding Geometry: Fusing Region and Depth for Highly Efficient 3D Tracking of Textureless Objects | IEEE Conference Publication | IEEE Xplore

作者提供代码：

3DObjectTracking/ICG at master · DLR-RM/3DObjectTracking · GitHub

1、依赖

The following dependencies are required: Eigen 3, GLEW, GLFW 3, and OpenCV 4. In addition, images from an Azure Kinect or RealSense camera can be streamed using the K4A and realsense2 libraries.

必须：

（1）Eigen 3：Eigen3+Ubuntu20.04安装_weixin_54470372的博客-CSDN博客

（2）GLEW：GLEW+Ubuntu20.04安装_weixin_54470372的博客-CSDN博客

（3）GLFW 3：GLFW+Ubuntu20.04安装_weixin_54470372的博客-CSDN博客

（4）OpenCV 4：OpenCV+Ubuntu20.04安装_weixin_54470372的博客-CSDN博客

非必须：

（5）kinect或realsense的库，本次未安装

（6）OpenMP，未安装

（7）Doxygen：Doxygen+Graphviz+Ubuntu20.04安装_weixin_54470372的博客-CSDN博客

（8）Graphviz：Doxygen+Graphviz+Ubuntu20.04安装_weixin_54470372的博客-CSDN博客

~/3dTracking$ git clone https://github.com/DLR-RM/3DObjectTracking.git
Cloning into '3DObjectTracking'...
remote: Enumerating objects: 325, done.
remote: Counting objects: 100% (325/325), done.
remote: Compressing objects: 100% (230/230), done.
remote: Total 325 (delta 159), reused 233 (delta 81), pack-reused 0
Receiving objects: 100% (325/325), 4.76 MiB | 1.49 MiB/s, done.
Resolving deltas: 100% (159/159), done.

2、源代码的文件结构

有5个文件夹：

• include/: header files of the ICG library

                                ICG库的头文件

• src/: source files of the ICG library

                                ICG库的源文件

• third_party/: external header-only libraries

                                只包含头文件的外部库

• examples/: example files for tracking as well as for evaluation on different datasets

                                用于跟踪和评估不同数据集的示例文件

• doc/: files for documentation

                                文档

3、CMAKE & BUILD

Use CMake to build the library from source. The following dependencies are required: Eigen 3, GLEW, GLFW 3, and OpenCV 4. In addition, images from an Azure Kinect or RealSense camera can be streamed using the K4A and realsense2 libraries. Both libraries are optional and can be disabled using the CMake flags USE_AZURE_KINECT, and USE_REALSENSE. If CMake finds OpenMP, the code is compiled using multithreading and vectorization for some functions. Finally, the documentation is built if Doxygen with dot is detected. Note that links to classes that are embedded in this readme only work in the generated documentation.

翻译：使用CMake从源代码构建库。需要以下依赖项:特征3、GLEW、GLFW 3和OpenCV 4。此外，来自Azure Kinect或RealSense相机的图像可以使用K4A和realsense2库进行流媒体传输。这两个库都是可选的，可以使用CMake标志USE_AZURE_KINECT和USE_REALSENSE来禁用。如果CMake找到了OpenMP，则使用多线程和对某些函数进行向量化来编译代码。最后，如果检测到Doxygen内嵌dot，则构建文档。注意，到这个自述文件中嵌入的类的链接只能在生成的文档中使用。

（1）cmake

首先，切换到目录3DObjectTracking/ICG目录，因为没有安装kinect和realsense的libraries，所以需要设置：USE_AZURE_KINECT和USE_REALSENSE两个options，所以输入命令：

cmake -DUSE_AZURE_KINECT=OFF -DUSE_REALSENSE=OFF

输出如下：

$ cmake -DUSE_AZURE_KINECT=OFF -DUSE_REALSENSE=OFF
CMake Warning:
  No source or binary directory provided.  Both will be assumed to be the
  same as the current working directory, but note that this warning will
  become a fatal error in future CMake releases.


CMake Warning (dev) at /usr/local/share/cmake-3.25/Modules/FindOpenGL.cmake:315 (message):
  Policy CMP0072 is not set: FindOpenGL prefers GLVND by default when
  available.  Run "cmake --help-policy CMP0072" for policy details.  Use the
  cmake_policy command to set the policy and suppress this warning.

  FindOpenGL found both a legacy GL library:

    OPENGL_gl_LIBRARY: /usr/lib/x86_64-linux-gnu/libGL.so

  and GLVND libraries for OpenGL and GLX:

    OPENGL_opengl_LIBRARY: /usr/lib/x86_64-linux-gnu/libOpenGL.so
    OPENGL_glx_LIBRARY: /usr/lib/x86_64-linux-gnu/libGLX.so

  OpenGL_GL_PREFERENCE has not been set to "GLVND" or "LEGACY", so for
  compatibility with CMake 3.10 and below the legacy GL library will be used.
Call Stack (most recent call first):
  CMakeLists.txt:22 (find_package)
This warning is for project developers.  Use -Wno-dev to suppress it.

-- Found Doxygen: /usr/bin/doxygen (found version "1.8.17") found components: doxygen dot 
-- Configuring done
-- Generating done
-- Build files have been written to: /home/r****/3dTracking/3DObjectTracking/ICG

（2）make

make

最后几行输出：

Patching output file 35/35
lookup cache used 1814/65536 hits=11252 misses=1888
finished...
[100%] Built target doc_doxygen

5、USAGE

As explained previously, ICG is a library that supports a wide variety of tracking scenarios. As a consequence, to start tracking, one has to first configure the tracker. For this, two options exist:

• One option is to use C++ programming to set up and configure all objects according to ones scenario. An example that allows running the tracker on a sequence streamed from an AzureKinect is shown in examples/run_on_camera_sequence.cpp. The executable thereby takes the path to a directory and names of multiple bodies. The directory has to contain Body and StaticDetector metafiles that are called .yaml file and _detector.yaml. Similarly, examples/run_on_recorded_sequence.cpp allows to run the tracker on a sequence that was recorded using record_camera_sequence.cpp. The executable allows the tracking of a single body that is detected using a ManualDetector. It requires the metafiles for a LoaderColorCamera, Body, and ManualDetector, as well as the path to a temporary directory in which generated model files are stored.

• In addition to the usage as a library in combination with C++ programming, the tracker can also be configured using a generator function together with a YAML file that defines the overall configuration. A detailed description on how to set up the YAML file is given in Generator Configfile. An example that shows how to use a generator is shown in examples/run_generated_tracker.cpp. The executable requires a YAML file that is parsed by the GenerateConfiguredTracker() function to generate a Tracker object. The main YAML file thereby defines how individual objects are combined and allows to specify YAML metafiles for individual components that do not use default parameters. An example of a YAML file is given in examples\generator_example\config.yaml.

翻译：

正如前面所解释的，ICG是一个支持多种跟踪场景的库。因此，要开始跟踪，首先必须配置跟踪器。对此，有两种选择:

• 一种选择是使用c++编程来根据一个场景设置和配置所有对象。示例/run_on_camera_sequence.cpp中显示了一个允许在来自AzureKinect的序列流上运行跟踪器的例子。因此，可执行文件采用一个目录的路径和多个主体的名称。该目录必须包含名为.yaml的Body和StaticDetector的meatafiles和_detect .yaml文件。类似地，examples/run_on_recorded_sequence.cpp允许在使用了record_camera_sequence.cpp进行记录的序列上运行跟踪器。可执行文件允许跟踪使用ManualDetector检测到的单个主体。它需要LoaderColorCamera、Body和ManualDetector的metafiles，以及存储生成的模型文件的临时目录的路径。
• 除了将跟踪器作为库与c++编程结合使用外，还可以使用生成器函数和定义总体配置的YAML文件来配置跟踪器。Generator Configfile中给出了关于如何设置YAML文件的详细描述。在examples/run_generated_tracker.cpp中显示了如何使用生成器的示例。可执行文件需要一个由GenerateConfiguredTracker()函数解析的YAML文件，以生成一个Tracker对象。因此，主YAML文件定义了如何组合单个对象，并允许为不使用默认参数的单个组件指定YAML元文件。在examples\generator_example\config.yaml中给出了一个YAML文件的例子。

本次不使用kinect或realsense，所以暂时不需要

6、评估

The code in examples/evaluate__dataset.cpp and examples/parameters_study_.cpp contains everything for the evaluation on the YCB-Video, OPT, Choi, and RBOT datasets. For the evaluation, please download the YCB-Video, OPT, Choi, or RBOT dataset and adjust the dataset_directory in the source code. Note that model files (e.g. 002_master_chef_can_depth_model.bin, 002_master_chef_can_region_model.bin, ...) will be created automatically and are stored in the specified external_directory. For the evaluation of the YCB-Video dataset, please unzip poses_ycb-video.zip and store its content in the respective external_directory. For the Choi dataset, the Matlab script in examples/dataset_converter/convert_choi_dataset.m has to be executed to convert .pcd files into .png images. Also, using a program such as MeshLab, all model files have to be converted from .ply to .obj files and stored in the folder external_directory/models. Both the OPT and RBOT datasets work without any manual changes.

翻译：

（1）examples/evaluate_ _dataset .cpp和examples/parameters_study_.cpp中的代码包含了对YCB-Video、OPT、Choi和RBOT数据集的所有评估。对于评估，请下载YCB-Video、OPT、Choi或RBOT数据集，并修改源代码中的dataset_directory。

（2）模型文件(如002_master_chef_can_depth_model.bin, 002_master_chef_can_region_model.bin，…)将自动创建，并存储在指定的external_directory中。

（3）数据集处理：对于YCB-Video数据集的评估，请解压缩poses_ycb-video.zip并将其内容存储在相应的external_目录中。对于Choi数据集，在examples/dataset_converter/convert_choi_dataset中的Matlab脚本。必须执行M将.pcd文件转换为.png图像。此外，使用像MeshLab这样的程序，所有的模型文件必须从.ply转换为.obj文件，并存储在external_directory/models文件夹中。OPT和RBOT数据集都不需要任何手动更改即可工作。

（1）数据集

（1）使用的数据集：YCB-Video, OPT, Choi, RBOT

（2）对于YCB-Video数据集的评估，请解压缩poses_ycb-video.zip并将其内容存储在相应的external_目录中。

（3）对于Choi数据集，在examples/dataset_converter/convert_choi_dataset中的Matlab脚本。必须执行M将.pcd文件转换为.png图像。

（4）OPT和RBOT数据集都不需要任何手动更改即可工作。

• YCB：PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes – UW Robotics and State Estimation Lab
• OPT：OPT Dataset
• Choi：Changhyun Choi > Research > RGB-D Tracking
• RBOT：Computer Vision and Mixed Reality Group | RBOT | Computer Vision and Mixed Reality Group

（2） 针对OPT数据集的评估

修改代码中的directory：

examples/evaluate_opt_dataset.cpp需要修改dataset、external和result_path

examples/parameters_study_opt_dataset.cpp需要修改dataset、external_path

dataset_path：存放数据集的目录

external_path：将要存放模型文件的目录

result_path：将要存放结果的目录

以下是我设置的目录：

  std::filesystem::path dataset_directory{"/home/r*/3dTracking/datasetP/OPT/Model3D/"};
  std::filesystem::path external_directory{"/home/r*/3dTracking/3DObjectTracking/ICG/examples/externalP/"};
  std::filesystem::path result_path{"/home/r*/3dTracking/3DObjectTracking/ICG/examples/resultP/"};