3D Gaussian Splatting

NeRF (Neural Radiance Fields) and its successor Gaussian Splatting can reconstruct three-dimensional scenes that can be freely navigated from any viewpoint from 2D images.

3D Gaussian Splatting (3DGS) was proposed in 2023 by the French INRIA team, achieving rendering speeds dozens of times faster than NeRF while maintaining comparable image quality.

3D Gaussian Splatting elevated neural rendering speed to real-time levels in 2023.

3DGS explicitly represents scenes using millions of three-dimensional Gaussian ellipsoids with attributes including position, covariance matrix, color (spherical harmonic coefficients), and opacity.

3D Gaussian Splatting uses explicit Gaussian spheres as basic units, achieving a substantial increase in rendering speed, and has rapidly gained adoption in real-time applications in recent years.

Compared to NeRF, 3DGS training is several to more than ten times faster, real-time rendering frame rates can reach tens to hundreds of FPS, and scenes can be directly exported as point cloud assets.

The neural rendering technical path, represented by NeRF (Neural Radiance Fields) and 3D Gaussian Splatting, has the potential to bypass the industrial pipeline of traditional polygon modeling.

The cost of 3D Gaussian splatting is higher VRAM usage, sensitivity of sorting and blending to overlapping regions, and susceptibility to artifacts at extreme viewing angles.

Unlike NeRF's volumetric rendering with dense sampling along rays, the rendering cost of 3D Gaussian splatting mainly depends on the number of Gaussian spheres and screen coverage area, so it can easily reach hundreds of frames per second.

3D Gaussian splatting is a real-time 3D reconstruction and rendering technique proposed in 2023 by institutions including INRIA.

Rendering in 3DGS is performed by projecting the Gaussians onto the image plane and compositing them by depth via differentiable alpha blending, enabling real-time frame rates.

3D Gaussian Splatting (3DGS) was proposed in 2023 and is an explicit 3D scene representation method that characterizes a scene with anisotropic 3D Gaussian ellipsoids, each Gaussian carrying position, covariance, opacity, and spherical harmonic coefficients.

3DGS uses rasterized splatting, in which mutual occlusion between different Gaussian primitives is not explicitly modeled.

VisibilitySplat proposes a method that retains the efficient training and visualization pipeline of rasterized 3DGS while additionally attaching a sparse visibility "oracle" to handle occlusion queries.

3D Gaussian Splatting (3DGS) is a 3D scene representation and rendering technology proposed in 2023, which explicitly represents scenes with millions to billions of 3D ellipsoids carrying direction, color, transparency and spatial scale, and achieves real-time rendering through rasterization.

3DGS depicts static geometry and appearance at a single moment and inherently contains no temporal dimension; making a scene move requires introducing time-varying deformation or an additional 4D representation.

Compared to NeRF, 3DGS renders tens of times faster and can directly output geometric information usable by physics engines.

SplatQ is a combinatorial optimization pruning algorithm for 3D Gaussian Splatting (3DGS) that coordinates regional budget allocation and joint Gaussian selection under a fixed budget.

Global Top-K pruning has two structural defects: lack of regional budget control and neglect of joint deletion effects.

Compared to NeRF, which requires dense ray-marching sampling, 3DGS uses a rasterization pipeline that is better suited for GPU parallelism and can achieve real-time frame rates.