Gaussian Splatting

I think I heard Sachin talk about this

But also https://twitter.com/Arata_Fukoe/status/1714931950719508967

This is really fundamental stuff that I need to learn alongside NeRF. It’s super cool.

I need to look into Gaussian SLAM.

Neural Surface Reconstruction

NeRF excels in producing high-quality, photorealistic renderings at the cost of computational efficiency, while Gaussian Splatting offers a faster and more memory-efficient alternative with potentially lower rendering quality. The choice between the two depends on the specific requirements of the application, such as the need for real-time rendering versus the demand for high-fidelity imagery.

Walkthrough (CS231n 2025 Lec 15)

Representation

Parameterize the scene as a sparse set of 3D Gaussian blobs (Kerbl et al. SIGGRAPH 2023). Each blob stores:

  • Mean (position)
  • Covariance (anisotropic extent, via scale + rotation)
  • Density / opacity
  • View-dependent color (spherical-harmonic coefficients)

Think of it as an explicit, extended point cloud — the opposite end of the spectrum from NeRF, which bakes the whole scene into a dense MLP.

Rendering

Project each Gaussian to screen space → alpha-blend front-to-back per pixel (tile-based rasterizer). No MLP in the inner loop — this is what unlocks real-time rendering.

Numbers (vs MipNeRF360)

3DGS (30K iters)MipNeRF360
SSIM0.830.80
PSNR26.9127.11
Train time38 min48 h
Render FPS1370.07

~2000× faster rendering, ~75× faster training, comparable quality. This is why 3DGS displaced NeRF for most practical capture-and-playback work starting 2024.

NeRF vs 3DGS framing (CS231n 2024 Lec 18, slide 97)

The clean one-liner from the 2024 deck:

NeRF3D Gaussian Splatting
Inner loopQuery a continuous MLP along the rayBlend a discrete set of Gaussians along the ray
FittingSeveral hours on best GPUsFew minutes
Rendering~10 s/frame at moderate resolutionReal-time

Variants (CS231n 2024 Lec 18)

  • Dynamic 3D Gaussians (Luiten et al. 3DV 2024) — track a moving scene over time by persisting Gaussians frame-to-frame; gives free dense correspondences across frames as a side effect.
  • Gaussian Splatting SLAM (Matsuki et al. CVPR 2024) — use a 3DGS scene as the dense map in a SLAM loop; simultaneously localize the camera and grow/refine Gaussians as new views arrive.