2D Gaussian Splatting for Surface Reconstruction

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2207 papers307 conceptsfrozen 2026-10-04

Standard 3D Gaussian Splatting (3DGS) revolutionized real-time radiance fields by representing scenes as a collection of volumetric "blobs." While these 3D ellipsoids are excellent at capturing view-dependent appearance for novel view synthesis, they are fundamentally unsuited for geometry. Because 3D Gaussians have volume but no explicit boundary, the resulting reconstructions often suffer from "floaters," jagged surfaces, and ill-defined normals.

To solve this, researchers have moved toward 2D Gaussian Splatting—also known as Gaussian Surfels. By flattening 3D volumes into oriented 2D disks, this approach bridges the gap between point-based rendering and continuous surface reconstruction.

From Volumetric Blobs to Planar Surfels

The core intuition behind 2D Gaussian Splatting, pioneered by Binbin Huang and Zehao Yu in 2D Gaussian Splatting for Geometrically Accurate Radiance Fields, is to replace the 3D covariance matrix with a 2D representation. Instead of a "fuzzy" ellipsoid, each primitive is a flat, oriented disk—a surfel.

This shift resolves the inherent ambiguity of surface location. In a volumetric model, a ray might pass through several overlapping blobs, making it difficult to determine exactly where the "surface" begins. In contrast, 2D Gaussians act as local tangent space elements. As Shuai Zhang notes in Dynamic 2D Gaussians, this forces the model to prioritize surface alignment over volumetric density, leading to cleaner, watertight meshes that naturally wrap around the geometry of the scene.

The Power of Precise Normals

The most significant advantage of 2D primitives is the derivation of mathematically rigorous surface normals. In 3DGS, normals are often "faked" or estimated from density gradients, which are notoriously noisy. With 2D Gaussians, the normal is explicitly defined by the cross-product of the disk's two tangent vectors.

Precise normals are the "north star" for several advanced rendering tasks:

  • Specular Reflections: High-frequency reflections are extremely sensitive to surface orientation. Yuanhu Yao and Li Zhang demonstrate in Reflective Gaussian Splatting that using 2D disks allows for accurate ray-tracing of mirror-like finishes that volumetric methods typically blur.
  • Inverse Rendering: To separate an object's true color (albedo) from the light hitting it, you must know how the surface is tilted relative to the light source. Georgios Kouros and Minye Wu utilize this in Spec-Gloss Surfels to achieve high-fidelity relighting of glossy objects.
  • Physical Interactions: For robotics, "cloudy" artifacts can lead to grasping failures. Yu Sheng shows in MSGField that 2D primitives minimize depth distortion, allowing robots to interact with thin objects like cable ties.

Consistency Across Time and Scale

The geometric constraint of 2D surfels also provides a powerful inductive bias for large-scale and dynamic scenes. In outdoor environments, Kele Shao found that 2D surfels prevent the "surface thickening" common in man-made structures like walls and floors, improving reconstruction accuracy by nearly 20% in LI-GS.

For moving objects, the challenge is maintaining a consistent "skin" as the object deforms. Jae‐Seong Lee and Marcel C. Buehler address this in SurFhead by using Jacobian Blend Skinning to deform 2D Gaussian disks. Because the primitives are thin, they can shear and stretch to follow complex facial expressions without the "popping" artifacts seen in volumetric point clouds.

Why It Works: The Differentiable Rendering Synergy

Ultimately, 2D Gaussian Splatting works because it combines the speed of rasterization with the geometric rigor of meshes. By using a near-exact differentiable rendering algorithm—as explored in Geometry Field Splatting with Gaussian Surfels—the system can optimize the position, orientation, and scale of these disks directly from images.

The result is a representation that is as fast as 3DGS (often reaching thousands of frames per second, as shown by Xinjie Zhang in GaussianImage) but provides the clean, relightable, and physically consistent surfaces required for the next generation of 3D vision.

Papers in this version

Showing 60 of 207 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 125 papers
  2. Version 1 · 2026-10-04 · 207 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_planar_primitive_representation_3726a51c,
  title = {planar primitive representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/planar-primitive-representation/rdr_3726a51c8ee44855942de8fa9d91221b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_2d_gaussian_splatting_for_surface_recons_3726a51c_v1,
  title = {2D Gaussian Splatting for Surface Reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/2d-gaussian-splatting-for-surface-reconstruction/rdr_3726a51c8ee44855942de8fa9d91221b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}