ImplicitFeatureEncoding

This repository implements the method described in the following paper:

Zexin Yang, Qin Ye, Jantien Stoter, and Liangliang Nan. 
Enriching Point Clouds with Implicit Representations for 3D Classification and Segmentation.
Remote Sensing. 15(1), 61, 2023.

Build

The current implementation depends on the Point Cloud Library (PCL). Please install PCL first.

To build ImplicitFeatureEncoding, you need CMake (>= 3.12) and a compiler that supports >= C++17. With CMake, ImplicitFeatureEncoding can be built on almost all platforms, although so far we have only tested it on Linux (GCC >= 4.8, Clang >= 3.3).

There are many options to build ImplicitFeatureEncoding. Choose one of the following (not an exhaustive list):

• Option 1 (purely on the command line): Use CMake to generate Makefiles and then make (on Linux/macOS) or nmake(on Windows with Microsoft Visual Studio).

  • On Linux or macOS, you can simply

    $ cd path-to-root-dir-of-ImplicitFeatureEncoding
    $ mkdir Release
    $ cd Release
    $ cmake -DCMAKE_BUILD_TYPE=Release ..
    $ make
    

  • On Windows with Microsoft Visual Studio, use the x64 Native Tools Command Prompt for VS XXXX (don't use the x86 one), then

    $ cd path-to-root-dir-of-ImplicitFeatureEncoding
    $ mkdir Release
    $ cd Release
    $ cmake -G "NMake Makefiles" -DCMAKE_BUILD_TYPE=Release ..
    $ nmake
    

• Option 2: Use any IDE that can directly handle CMakeLists files to open the CMakeLists.txt in the root directory of ImplicitFeatureEncoding. Then you should have obtained a usable project and just build it. I personally highly recommend using CLion. For Windows users: your IDE must be set for x64.

• Option 3: Use CMake-Gui to generate project files for your IDE. Then load the project to your IDE and build it. For Windows users: your IDE must be set for x64.

Don't have any experience with C/C++ programming? Have a look at Liangliang's step-by-step tutorial.

Usage

The current implementation directly generates augmented point clouds, i.e., xyz (+ rgb) + implicit features. These augmented point clouds can then be used as input for point-based neural networks. The main.cpp file demonstrates how to use the ImplicitFeatureEncoding class. To encode implicit features, run the built executable like this:

./ImplicitFeatureEncoding  <input_path>  <sample_sphere_ply_file>  <neighbor_radius>  <whether_to_fix_z_axis>  <output_path>

• <input_path> path to the input point cloud files (ply format)
• <sample_sphere_ply_file> path to the sample sphere file (ply format)
• <neighbor_radius> radius used for spherical neighborhood search
• <whether_to_fix_z_axis> set to 1 for real-world scene datasets (e.g., S3DIS and SensatUrban, and set to 0 for synthetic datasets (e.g., ModelNet40)
• <output_path> path where the resulting augmented point cloud files will be saved

Citation

We kindly ask you to cite our paper if you use (part of) the code or ideas in your academic work:

@article{yang2022enriching,
  title={Enriching Point Clouds with Implicit Representations for 3D Classification and Segmentation},
  author={Yang, Zexin and Ye, Qin and Stoter, Jantien and Nan, Liangliang},
  journal={Remote Sensing},
  volume={15},
  number={1},
  pages={61},
  year={2022},
  publisher={MDPI}
}

---

Please feel free to contact me at zexinyang@gzpi.com.cn with questions, comments, or suggestions ;-)

Zexin Yang

December 12, 2022