Gaussian Splat
A 3D rendering technique that represents a scene as millions of small fuzzy ellipsoids ("Gaussians") instead of polygon meshes. Each splat has position, colour and shape; rendering blends them together for a photorealistic view.
Also known as: 3D Gaussian Splatting, Gaussian splatting
Gaussian splatting is a 3D scene-representation technique that emerged from the SIGGRAPH 2023 “3D Gaussian Splatting for Real-Time Radiance Field Rendering” paper. It replaces the polygon-mesh approach that dominates traditional 3D graphics with a cloud of overlapping anisotropic Gaussians. Each Gaussian is described by a position, an orientation, a scale, a colour and an opacity. Rendering sorts the splats by distance from the camera and composites them in order, which produces photorealistic novel-view synthesis at real-time frame rates without the data-collection overhead of building a full mesh.
The advantage over earlier neural rendering approaches (notably Neural Radiance Fields, or NeRF) is speed and editability. A NeRF requires a neural network forward pass per pixel and is slow to train and render; a Gaussian splat is a flat parameter set that GPUs render in roughly the same way they handle particle systems. Once trained, a splat can be edited, scaled, sliced and composited into other splats much more easily than a NeRF can. For applications that need a 3D map of a physical space that loads quickly and renders well on consumer hardware, splatting has displaced NeRF as the default.
In the spatial-AI stack used by Auki’s posemesh, reconstruction servers consume scans from the Domain Management Tool and produce either point clouds (sparse, fast to index) or Gaussian splats (dense, photoreal, navigable). The splat output is what makes a stored 3D map of a retail store useful for an in-app or smart-glasses view of inventory; the point-cloud output is what feeds the position-and-orientation reasoning that lets a customer’s phone know which aisle it is in.
The data-sovereignty implication for DeAI is that splat data of a physical space encodes a lot more than a flat image: layout, dimensions, occupancy patterns and (if humans were present during capture) identifiable movement traces. Whether a decentralised perception network keeps splat data on the venue owner’s domain server or pushes it to centralised reconstruction infrastructure is a meaningful privacy choice.