Gaussian Splatting, a volumetric technique for Real-Time 3D Rendering
Gaussian Splatting is a novel1 technique for real-time 3D rendering, specifically designed for photogrammetry generated meshes.
What is it for?
Faster, accurate, dense 3D point clouds, which are a byproduct of photogrammetry.
How does it work?
It is based on volume rendering techniques that deals with the direct rendering of volume data without converting the data into surface or line primitives. It is an iterative process that continuously perturbs geometric representations from previous passes and renders these intermediate shapes. The core idea relies on Gaussian kernels applied over areas, moving from one point to another. Gaussian distribution fits these spaces with mathematical weightings that dictate each influence or weight of each kernel parameter in calculating points by point rendering.
Stages
- Input: 3D points and desired interpolation characteristics are loaded as input for the CPU or GPU.
- Geometry preparation: Geometric information is extrapolated to generate intermediate geometric structure components from Gaussian Splatting, such as sphere types or pre-sampled convex primitives.
- Decimation: Kernels are deployed on the surface, giving area representative characteristics from the points.
Benefits of Gaussian Splatting
- Fast(er) rendering: Gaussian Splatting allows for fast rendering of complex scenes, making it ideal for real-time applications.
- High-quality visuals: The technique produces high-quality visuals with convincing geometry and texture mapping.
- Efficient memory usage: Gaussian Splatting uses less memory than traditional techniques, making it suitable for devices with limited memory.
Limitations of Gaussian Splatting
- Computational complexity: The technique requires significant computational resources, which can be a challenge for devices with limited processing power.
- Optimization: May require careful optimization to achieve the best results, which can be time-consuming and require significant expertise.
- Scalability: Gaussian Splatting may not be suitable for very large scenes or complex models, due to its computational requirements.
Additional Technical Limitations
- Kernel selection: Choosing the optimal kernel for a given scene can be challenging, and requires careful consideration of factors such as kernel size, decay rate, and slope.
- Sampling density: The sampling density of the point cloud can affect the quality of the rendered image, and may require careful adjustment to achieve the best results.
- Control parameters: The technique requires careful adjustment of control parameters, such as the number of iterations, kernel size, and decay rate, to achieve the best results.