This project contains a C++ OpenGL viewer for 3D gaussian splatting scenes.
Pre-trained models can be downloaded from: "https://github.com/graphdeco-inria/gaussian-splatting/tree/main"
The main particularity is that the gaussians are rendered with standard alpha blending instead of the custom tile-based renderer used in the 3DGS code.
Note: An nvidia GPU supporting OpenGL 4.6 is required, with the cuda toolkit installed.
The rendering loop is as follows:
- Cull the gaussians outside of the view frustum (see resources/shaders/testVisibility.cp)
- Sort the gaussians either front to back or back to front (with cub's radix sort and opengl/cuda interop), see src/Sort.cu
- Compute oriented bounding boxes for the remaining gaussians (see resources/shaders/computeBoundingBoxes.cp)
- Evaluate the spherical harmonic basis of each gaussian (resources/shaders/predict_colors.cp)
- Render the gaussians with the hardware rasterizer and alpha blending (see resources/shaders/quad.vs and resources/shaders/quad.fs.
The main rendering loop is in src/GaussianCloud.cpp.
The opengl shaders are located in the resources/shaders directory. The file extensions are .cp for compute shaders, .vs for vertex shaders and .fs for fragment shaders. GLSL header files are located in resources/shaders/common. The shaders use special comments (/*-- and --*/ ) so that they can be edited with a standard C++ IDE. These comments are remove at runtime with a regex to produce valid GLSL code.
Note: The viewer will search for a file named "bicycle.ply" in the current directory when clicking "Load ply" (see src/Window.cu:239 to change the file name).
- Step 1:
Install glfw. Open a developper command prompt ("Developper PowerShell for VS 2022").
./install_script_win.ps1
- Step 2: Build the program. Make sure to adapt the command line to your specific cuda version, paths, etc...
$install_dir = (Join-Path $pwd "/libs_install") mkdir build cd build cmake .. -DCMAKE_PREFIX_PATH="$install_dir" -DCMAKE_GENERATOR_TOOLSET="cuda=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.6" cmake --build . --target INSTALL --config Release -j cd ../
cp ./libs_install/bin/glfw3.dll ./build/Release/
Run from root directory with:
./build/Release/HardwareRasterized3DGS.exe