4D-Rotor Gaussian Splatting: Towards Efficient Novel View Synthesis for Dynamic Scenes

Duan et al. introduce 4D-Rotor GaussianSplatting (4DRotorGS), a dynamic extension of 3D Gaussian Splatting that represents a time-varying scene with anisotropic 4D Gaussians defined over the joint XYZT domain. Prior dynamic NeRF and dynamic-Gaussian work generally encodes motion via a canonical 3D representation plus an implicit or explicit DeformationField that maps each timestamp back to canonical coordinates. Such canonical formulations struggle with abrupt motions, topological changes, and emerging or disappearing content, because the deformation network must absorb discontinuities that violate its smoothness assumptions.

4DRotorGS avoids the canonical-plus-deformation factorisation entirely. Each primitive is a full 4D Gaussian with an orientation parameterised by a RotorRepresentation drawn from geometric algebra, which generalises the quaternion approach used for 3D rotations to four dimensions while remaining numerically stable. At any rendering time t the authors take a temporal slice through the 4D field, producing a set of effective 3D Gaussians for that frame that can be splatted with the standard differentiable rasteriser. This temporal slicing naturally handles abrupt motion: parts of the scene that appear or disappear are represented by 4D Gaussians with narrow temporal extent, rather than by a smooth deformation of a persistent canonical primitive.

The authors implement temporal slicing and splatting in a CUDA pipeline optimised alongside the existing 3DGS kernels, achieving 277 FPS at the benchmark resolution on an RTX 3090 and 583 FPS on an RTX 4090. RealTimeRendering performance and reconstruction quality are competitive with or better than prior dynamic methods, including the contemporary 4D Gaussian Splatting (4D-GS) line of work that uses a HexPlane-style deformation field instead.