What is Gaussian splatting?
Gaussian splatting represents a 3D scene explicitly, as a large collection of tiny colored, semi-transparent 3D blobs called Gaussians. Reconstructed from a set of images, these Gaussians can be rendered from any viewpoint by projecting and blending them onto the image plane, a step that is fast enough for real-time, interactive viewing.
It achieves NeRF-like photorealism while sidestepping the slow rendering of implicit neural representations.
Key takeaways
- It represents a scene as many colored 3D Gaussians.
- It reconstructs from images and renders in real time.
- It offers a faster, explicit alternative to NeRF.
How it works
Starting from images with known camera poses, an optimization places, colors, and shapes many Gaussians so that, when projected and blended, they reproduce the input views. Because rendering is a direct projection of explicit primitives rather than sampling a neural network along rays, it runs far faster than NeRF, enabling smooth interactive navigation of the reconstructed scene.
Why it matters
Real-time, photorealistic 3D from ordinary images is valuable for building simulation environments, digital assets, and immersive views, all relevant to physical AI and content creation. Gaussian splatting's speed made 3D reconstruction practical for interactive use, complementing the point clouds and depth maps used elsewhere in 3D perception.
Frequently asked questions
How is Gaussian splatting different from NeRF?
Both reconstruct 3D scenes from images, but Gaussian splatting uses explicit Gaussians that render in real time, while NeRF uses an implicit neural network that is slower to render.
What is a Gaussian in this context?
A small, colored, semi-transparent 3D blob whose position, shape, and color are optimized so many of them together represent the scene.
Related terms