3DGS Viewer Logo
3DGS Viewer
Back to guide
English中文

Research Paper

2D Gaussian Splatting for Geometrically Accurate Radiance Fields

A geometry-focused paper that replaces volumetric 3D Gaussians with oriented 2D Gaussian disks to better model surfaces.

March 2024GeometryarXiv:2403.17888
arXiv Paper->

Detailed Reading

2DGS changes the primitive itself. Vanilla 3DGS uses ellipsoids, which are excellent for radiance but can blur across surfaces or represent thickness that does not exist. 2DGS instead represents surfaces with oriented Gaussian disks embedded in 3D space.

The rendering algorithm computes ray-disk intersections and splats the resulting contributions with perspective correctness. The training objective adds depth distortion and normal consistency terms, pushing the model toward surfaces that agree across views rather than volumetric blobs that merely match colors.

The paper is a major geometry correction to the original 3DGS idea. It says that if the goal is accurate surfaces, the representation should encode surfaces directly. This is why 2DGS belongs beside SuGaR and GOF in any serious reading list.

2D Gaussian Splatting changes the primitive to better match surfaces. Standard 3D Gaussians are ellipsoids with volume, which can look good but creates ambiguous geometry. This paper represents local surface elements as oriented 2D Gaussian disks, making the optimized primitives more naturally surface-aligned.

The renderer still splats projected Gaussians, but the mathematical model is tied to a disk-like support rather than a full volumetric ellipsoid. That makes normals and depth behavior more meaningful for surface reconstruction. It also reduces the tendency for thick clouds to explain a thin physical surface.

The algorithmic point to watch is how geometry and appearance supervision are balanced. The model needs enough flexibility to match images, but enough constraint to keep disks on a surface. Losses and regularizers are designed so the representation can be converted into accurate geometry rather than only render nice images.

The paper is important for anyone who wants meshes, collision, measurement, or CAD-like downstream use. It may be less forgiving than volumetric 3DGS in fuzzy, transparent, or poorly observed areas, but it advances the idea that splatting can be a geometry representation, not just a radiance-field renderer.

What The Paper Does

2D Gaussian Splatting argues that 3D Gaussians are not ideal for surface geometry because they are volumetric and can be multi-view inconsistent.

The paper collapses the representation into oriented planar Gaussian disks, then uses perspective-correct splatting, depth distortion, and normal consistency losses to improve geometry.

Core Ideas

  • Models surfaces intrinsically with 2D Gaussian disks embedded in 3D.
  • Uses ray-splat intersection for perspective-accurate rasterization.
  • Adds geometric regularization for cleaner depth and normal reconstruction.

Why It Matters

  • It is one of the most important geometry-oriented alternatives to vanilla 3DGS.
  • It shows that changing the primitive shape can improve surface accuracy, not just adding losses.
  • It is central reading for anyone comparing splat rendering with mesh or surface reconstruction goals.

Read This If

  • You care about accurate geometry, normals, depth, or mesh extraction.
  • You are deciding whether 3D ellipsoids are the right primitive for your task.
  • You need a bridge between radiance-field quality and surface reconstruction quality.

Limitations And Caveats

  • 2D disks can be better for surfaces but are a different representation from standard 3DGS files.
  • The method targets geometry accuracy, so format/tool compatibility may require custom pipelines.
  • Very complex transparency or volumetric effects are still outside the core surface assumption.