3D Gaussian Splatting

A minimalist CUDA and C++ implementation of 3D Gaussian Splatting for real-time radiance field rendering. This repository provides a lightweight foundation for experimenting with the core ideas from the original paper.

Installing

1. Core Build Environment

You will need a standard C++ development toolchain. Please install the following using your system's package manager (e.g., apt for Debian/Ubuntu, brew for macOS).

• C++ Compiler with C++23 support (e.g., GCC 12+, Clang 15+)

• CMake (version 3.20 or later)

• OpenCV

For example, on Ubuntu 22.04, you can install these with:

sudo apt install cmake build-essential g++-12 libopencv-dev

2. NVIDIA CUDA Toolkit

A compatible NVIDIA driver and the CUDA Toolkit are required for the GPU-accelerated components.

• NVIDIA CUDA Toolkit (version 13.0 or later)

Please download and install the appropriate version for your system from the NVIDIA CUDA Toolkit website. After installation, ensure that the PATH, CPLUS_INCLUDE_PATH, and LD_LIBRARY_PATH environment variables are configured correctly in your shell profile (e.g., .bashrc).

Building

1. Clone the repository:

git clone --recursive https://github.com/AndrewBoessen/3DGS.git
cd 3DGS

2. Configure with CMake: Create a build directory and run CMake

cmake -S . -B build

Note: If you have multiple compilers, you can specify one, e.g., -DCMAKE_CXX_COMPILER=g++-12.

3. Compile the code: Build the project using CMake

cmake --build build --parallel $(nproc)

The main executable, gaussian_splatting, will be located in the build/ directory.

Using

Mip-NeRF 360

To train a scene from the Mip-NeRF 360 dataset specify the scene to train in the config file: base.yaml or extended.yaml and run the main executable.

./build/gaussian_splatting <path_to_config> <path_to_dataset>

For example

./build/gaussian_splatting ./config/extended.yaml ../Downloads/360_v2

Colab

A Colab notebook is available to train a demo scene and view rendered images.

COLMAP

Testing

The project includes a suite of unit tests built with Google Test. The tests are compiled automatically during the main build process.

To run all tests, execute ctest from within the build directory. The --output-on-failure flag is recommended to only show logs for failing tests.

cd build
ctest --output-on-failure

This will discover and run all test executables, including those for the CUDA kernels.