Gaussian Point Splatting
5 hours ago
- #GPU rendering
- #Parallel computing
- #Point sampling
- Introduces Gaussian point splatting, a stochastic rendering method for Gaussian splats that efficiently handles scenes with numerous Gaussians.
- Uses pixel-sized opaque points sampled from Gaussians and splats them using 64-bit atomics in a framebuffer.
- Achieves workload distribution across millions of threads via parallel programming primitives, though independent splatting can cause multiple points per pixel.
- Addresses challenges in determining point count per Gaussian and distribution for desired opacity, ensuring fidelity to original Gaussian splatting.
- Incorporates hierarchical frustum and occlusion culling to accelerate rendering, achieving real-time performance with hundreds of millions of Gaussians.
- Results differ slightly from original Gaussian splatting in noise and aliasing, while maintaining real-time scalability.