In 3D Gaussian splatting, a scene is represented by many overlapping elements whose contributions are projected and combined to form an image from a chosen viewpoint. Parameters can be fitted from captured views to approximate the scene's appearance.
The representation can support efficient rendering, but visual plausibility does not guarantee accurate geometry or physical behavior. Coverage, moving objects and viewpoints unlike the captured observations can expose artifacts or missing information.
ELI5
Gaussian splatting builds a view from many soft, overlapping colored patches placed in space. Together, the patches can make a scene look detailed when viewed from different positions.
For example, a collection of photographs could be used to fit a representation of a room that you can look around. Areas hidden from all the photographs may still be missing or look wrong.
Is a splatted scene a complete physics simulation?
No. An appearance representation does not by itself establish physical rules or reliable behavior under actions.
Does realistic rendering prove accurate geometry?
No. A scene can look convincing while containing spatial errors or artifacts.
