Planar Gaussian Primitives for Surface Reconstruction

227 concepts119 papers2 proposals

The advent of 3D Gaussian Splatting revolutionized real-time novel view synthesis by representing scenes as collections of millions of 3D ellipsoids. However, these ellipsoids are fundamentally volumetric "blobs." When tasked with reconstructing thin surfaces, sharp edges, or flat man-made structures, unconstrained 3D ellipsoids struggle. They tend to create floating artifacts in empty space, suffer from multi-view inconsistency, and lack a well-defined surface normal.

To bridge the gap between fast volumetric rendering and high-fidelity geometry, researchers have shifted toward planar primitive representations. By flattening 3D kernels into 2D disks, or "surfels," this paradigm enforces a strong surface prior directly within the radiance field.

From Volumetric Blobs to Planar Surfels

In standard 3D Gaussian Splatting, each primitive possesses three scaling parameters defining its extension in 3D space. As explored by Danpeng Chen et al. in PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction, constraining one of these scaling parameters to be extremely thin effectively transforms the primitive into a 2D planar disk.

This geometric constraint fundamentally alters how the optimization landscape behaves. Instead of filling space with unstructured volumetric clouds, the primitives are forced to align tightly with the actual physical skin of the object.

Standard 3D Gaussian           Planar 2D Surfel
     (Volumetric)                 (Surface-Aligned)
       _______                         _______
      /       \                       /       \
     |    •    |                     |----•----|  <-- Explicit Normal
      \_______/                       \_______/

To prevent these flattened primitives from degenerating into one-dimensional "needles" during optimization, researchers must carefully manage their topology. Junha Hyung et al. address this in Effective Rank Analysis and Regularization for Enhanced 3D Gaussian Splatting by analyzing the effective rank of the primitives' shape matrices. By penalizing rank-1 (needle-like) structures and encouraging rank-2 (disk-like) configurations, the optimization converges on smooth, continuous surfaces, cutting geometric reconstruction errors nearly in half.

Why Planar Primitives Work: The Power of Explicit Normals

The core reason planar primitives outperform volumetric blobs lies in their mathematical clarity. A 3D ellipsoid does not have a single, well-defined surface normal; its density fades out symmetrically in all directions. A 2D disk, however, possesses an explicit orientation vector.

This mathematical precision unlocks two critical capabilities:

1. Physically Grounded Light Transport

In tasks like inverse rendering, relighting, and specular reflection modeling, knowing the exact orientation of a surface is non-negotiable. Specular reflections rely on precise ray-surface intersection calculations. If the surface normals are fuzzy or noisy, reflected light rays scatter erratically, resulting in blurred or missing details.

As demonstrated by Jiachen Tao et al. in TraceFlow: Dynamic 3D Reconstruction of Specular Scenes Driven by Ray Tracing, utilizing 2D planar primitives provides the mathematically precise normals required to drive a hybrid ray-tracing pipeline. This allows the system to render sharp, moving reflections—such as an object moving across a curved mirror—that volumetric methods blur entirely.

Similarly, Yuxuan Yao et al. show in Reflective Gaussian Splatting that replacing 3D ellipsoids with 2D disks eliminates multi-view inconsistencies on flat surfaces, providing the clean normal fields necessary for physically based deferred shading.

2. Stable Mesh Extraction

Extracting a clean triangle mesh from a standard 3D Gaussian field is notoriously difficult because the volumetric blobs overlap and float arbitrarily. Planar primitives solve this by providing a consistent intersection plane across multiple viewpoints.

In MeshSplat: Generalizable Sparse-View Surface Reconstruction via Gaussian Splatting, Hanzhi Chang explains that this inherent flatness ensures the predicted geometry remains structurally valid even during view extrapolation. Because the primitives act as oriented local surface patches, downstream algorithms like Poisson surface reconstruction can seamlessly bind them into continuous, watertight manifold meshes suitable for physics engines, as shown by Decai Chen et al. in Adaptive and Temporally Consistent Gaussian Surfels for Multi-View Dynamic Reconstruction.

Representing Thin and Complex Topologies

Planar primitives are uniquely efficient at modeling thin structures that lack a defined interior volume. Representing a flat wall, a leaf, or a piece of fabric with round 3D kernels requires a massive number of overlapping spheres to avoid a "bumpy" appearance, leading to high memory overhead.

Shujuan Li leverages this efficiency in GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting, showing that "overfitting" flat 2D Gaussians onto complex, open boundaries allows for the rapid reconstruction of highly detailed topologies like clothing and foliage.

This surface-aligned nature also provides a powerful regularizer for dynamic scenes and tracking systems. In 4DTAM: Non-Rigid Tracking and Mapping via Dynamic Surface Gaussians, Hidenobu Matsuki et al. utilize 2D Gaussian disks to track non-rigid, deforming objects from a single RGB-D stream. Because each disk has an explicit orientation, the system can apply geometric regularization directly along the surface tangent, preventing the representation from tearing apart or clouding up during complex deformations.

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 2, frozen text):

@misc{lacuna_planar_primitive_representation_3726a51c_v2,
  title = {planar primitive representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/planar-primitive-representation/rdr_3726a51c8ee44855942de8fa9d91221b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake