3D Gaussian Splatting

A real-time radiance field technique that replaces traditional voxels with millions of learned 3D Gaussians for photorealistic rendering at 100+ FPS.

3D Gaussian Splatting (3DGS) represents a paradigm shift from NeRF by ditching costly neural network queries for a rasterization-based approach. The system initializes a point cloud from SfM data (Structure-from-Motion) and optimizes 3D Gaussians (defined by position, covariance, color, and alpha) to fit the scene. By using a tile-based sorter and a custom CUDA kernel, it achieves sub-millisecond rendering speeds on consumer hardware like an RTX 3080. This allows for instant navigation of complex environments with high-frequency details (reflections and thin structures) that traditional meshes often struggle to resolve.

https://repo-elmo.github.io/guassian_splatting/