Planar & Mirrored Gaussians

Updated 30 June 2026

Planar/Mirrored Gaussians are specialized Gaussian primitives that enforce local planarity and mirror symmetry in 3D scenes.
The methods optimize surface reconstruction using degenerated covariance structures and regularization techniques to enhance depth and normal estimation.
Mirrored Gaussians employ explicit Householder reflections to achieve physically accurate reflective rendering and symmetry-aware scene composition.

Planar and mirrored Gaussians are specialized forms of Gaussian primitives that play an essential role in contemporary computer vision, graphics, and geometric learning, particularly in frameworks for 3D reconstruction, scene rendering, and symmetry modeling. Planar Gaussians refer to degenerate ellipsoidal Gaussians whose support is constrained to an infinitesimally thin surface, typically enforcing or representing local planarity in a scene. Mirrored Gaussians, in contrast, refer to Gaussians that are constructed as explicit geometric reflections of originals across a symmetry plane—enabling physically consistent rendering of phenomena such as mirror or glass reflections. These primitives underpin a range of recent high-fidelity methods for surface reconstruction, scene editing, reflective rendering, and isoperimetric analysis.

1. Mathematical Formulation and Definition

Planar Gaussians are constructed by degenerating a 3D anisotropic Gaussian to have rank-2 covariance: Given $\mu\in\mathbb{R}^3$ (mean) and $\Sigma\in\mathbb{R}^{3\times 3}$ (covariance), the planar form is obtained by penalizing the smallest eigenvalue of $\Sigma$ , effectively collapsing the kernel onto a local tangent plane. The density becomes

$G(x) = \exp\left(-\frac{1}{2}(x-\mu)^\top \Sigma^{-1}(x-\mu)\right)$

with $\Sigma$ constrained such that $\min(\text{eig}(\Sigma))\to 0$ (Chen et al., 2024, Wu et al., 21 Nov 2025). In screen-space or for rendering, the primitive is further projected onto the image plane or a local tangent patch, using means $\mu^{(2D)}\in\mathbb{R}^2$ and a $2\times 2$ covariance.

Mirrored Gaussians are generated using plane reflections. Given a plane $\mathcal{P}: n^\top x + b = 0$ , the Householder reflection yields the mirrored mean and covariance as

$\hat\mu = R\mu + t,\qquad \hat\Sigma = R\Sigma R^\top$

where $\Sigma\in\mathbb{R}^{3\times 3}$ 0 and $\Sigma\in\mathbb{R}^{3\times 3}$ 1 (Liu et al., 2024, Liu et al., 17 Nov 2025). This linear mapping preserves the ellipsoidal structure and aligns principal axes as required by the scene symmetry.

2. Formation, Optimization, and Planar Constraints

Planar Gaussians are typically derived via explicit regularization of 3D Gaussians:

Scale Loss: SPAGS and GigaGS apply a loss penalizing the minimum eigenvalue/axis length of the covariance,

$\Sigma\in\mathbb{R}^{3\times 3}$ 2

enforcing the planar constraint by flattening the ellipsoid along its shortest axis (Wu et al., 21 Nov 2025, Chen et al., 2024).

Normal Extraction: The direction of the collapsed axis yields an explicit surface normal for each primitive, supporting accurate normal and depth estimation with alpha-composited rendering (Wu et al., 21 Nov 2025).
Surface Regularization: Additional local and global constraints (e.g., depth smoothness, Atlanta-world plane priors) enforce coherence and encourage alignment of Gaussians with floor, wall, or ceiling structures (Zhang et al., 29 Oct 2025).

Flattening reduces memory and compute: planar Gaussians can be stored as two axis-lengths, a 3D center, and a normal vector, reducing parameter count and supporting efficient large-scale reconstruction (Chen et al., 2024).

3. Mirrored Gaussians, Reflection Modeling, and Symmetry

Mirrored (or dual) Gaussians are directly instantiated to encode optical reflection:

Plane Reflection: The mean $\Sigma\in\mathbb{R}^{3\times 3}$ 3, covariance $\Sigma\in\mathbb{R}^{3\times 3}$ 4, and local frame are transformed via the Householder operator, creating a physically consistent virtual Gaussian (Liu et al., 17 Nov 2025, Liu et al., 2024).
View-Dependent Color: For reflected Gaussians, view-direction-dependent color coefficients (e.g., spherical harmonics) must be evaluated with the reflected viewer direction, ensuring radiometric correctness (Liu et al., 17 Nov 2025, Liu et al., 2024).
Dual-Rendering: Scene rendering involves two sets: the original (transmission) Gaussians and their reflected (mirror) counterparts. Accumulation and alpha-splatting are performed independently, then blended using mask and Fresnel terms for physically plausible composition (Liu et al., 17 Nov 2025, Liu et al., 2024).
Optimization: End-to-end losses include photometric error, mask alignment, covariance/scale regularization, plane distance losses, and constraints ensuring geometric and radiometric consistency across both the real and mirror scene representations (Liu et al., 2024, Liu et al., 17 Nov 2025).

This reflection framework is essential for high-fidelity rendering of scenes with mirrors or transparent planar media (e.g., glass).

4. Applications in Surface Reconstruction and Reflective Rendering

Planar and mirrored Gaussians underpin state-of-the-art methods for:

Surface Reconstruction: SPAGS uses planar Gaussians for articulated-object reconstruction from sparse images, leveraging planarization to obtain accurate normals and depth (Wu et al., 21 Nov 2025). GigaGS scales this to very large scenes by partitioning LoD-hierarchical planar Gaussians across distributed memory, enabling gigascale reconstruction (Chen et al., 2024). AtlasGS incorporates Atlanta-world planar priors via learnable plane indicators and explicit plane-based regularization, yielding globally consistent urban and indoor surfaces (Zhang et al., 29 Oct 2025).
Mirror and Glass Reflection: MirrorGaussian and TR-Gaussians directly model scene symmetry and planar transmission/reflection via explicit mirrored primitive construction, dual compositing, and Fresnel-weighted blending, achieving real-time photorealistic rendering with physically accurate appearance of reflective and transmissive media (Liu et al., 2024, Liu et al., 17 Nov 2025).

Representative results demonstrate substantial improvements over previous approaches in reconstruction fidelity, geometric consistency, and frame rates, including real-time performance and gigascale scalability.

5. Isoperimetric Analysis of Planar Mirrored Gaussians

Theoretical work on mirrored Gaussians in the plane analyzes isoperimetric problems for densities comprising sums of symmetric Gaussians:

For sum-of-Gaussians densities $\Sigma\in\mathbb{R}^{3\times 3}$ 5 with centers mirrored about an axis, the isoperimetric regions are conjectured—and supported by variational analysis—to be half-spaces bounded by vertical lines orthogonal to the axis of symmetry (Berry et al., 2016).
First variation and curvature analysis shows that vertical boundaries uniquely minimize weighted perimeter for given weighted area, compared to horizontal or arbitrary linear separators, under the mirrored Gaussian density (Berry et al., 2016).
In one dimension, the isoperimetric sets for double-Gaussian line densities are intervals (rays), reinforcing the conclusion that mirror symmetry necessitates axis-orthogonal optimal boundaries (Berry et al., 2016).

This theoretical framework provides justification for the use of mirrored planar Gaussians in enforcing geometric regularity and symmetry in computational representations.

6. Implementation and Computational Considerations

Implementations of planar and mirrored Gaussians employ:

Differentiable Splat Rendering: Both direct 2D alpha-splatting and EWA-filtering in screen space are used, typically in raster-shaders or CUDA kernels with custom backward passes (Liu et al., 2024, Liu et al., 17 Nov 2025).
Hierarchical and Partitioned Processing: Planarization enables efficient storage (reduction to two widths and a normal per primitive) and partitioning in hierarchical (LoD) structures, allowing distributed training and inference with $\Sigma\in\mathbb{R}^{3\times 3}$ 6+ primitives (Chen et al., 2024).
Implicit Structure and Semantic Guidance: Implicit-structured (MLP-decoded) approaches align planar Gaussians with planes predicted via Atlanta-world or semantic indicators, merging local smoothness and global structural regularity (Zhang et al., 29 Oct 2025).
Fine-to-Coarse and Few-Shot Refinement: Reconstruction methods such as SPAGS adopt a multi-stage procedure with initial coarse optimization (with planar and depth constraints), followed by diffusion-based or multi-view refinement incorporating surface consistency and view alignment (Wu et al., 21 Nov 2025).

7. Comparative Performance and Extensions

Empirical results demonstrate the advantages of planar and mirrored Gaussian approaches:

Significant gains in surface accuracy, completeness, and F1 scores are reported across synthetic and real-world benchmarks, with AtlasGS and GigaGS outperforming both conventional and prior Gaussian-based methods on ScanNet, Replica, and large-scale urban datasets (Zhang et al., 29 Oct 2025, Chen et al., 2024).
MirrorGaussian and TR-Gaussians achieve state-of-the-art SSIM, PSNR, and real-time rendering speeds on multiple mirror-centric scene datasets (Liu et al., 2024, Liu et al., 17 Nov 2025).
A plausible implication is that the explicit handling of planarity and symmetry not only improves reconstruction and rendering realism but also significantly enhances computational tractability and scalability for large scenes.

The framework is extensible: mirrored primitives can be incorporated into any planar-Gaussian-based scene representation with negligible overhead, enabling modeling of global or local symmetries in complex environments (Chen et al., 2024).

The systematic development and deployment of planar and mirrored Gaussians have transformed both theoretical understanding and practical capabilities for 3D reconstruction, large-scale scene modeling, and reflective rendering. Their mathematical construction leverages minimal geometric priors but facilitates interpretable, scalable, and physically accurate scene representations compatible with modern differentiable rendering and learning frameworks.